diff --git a/.gitignore b/.gitignore
index 05436bd..83b4611 100644
--- a/.gitignore
+++ b/.gitignore
@@ -20,3 +20,4 @@ node_modules/
 /composeApp/release/
 /test_books/
 /.junie/
+/.local/
diff --git a/README.md b/README.md
index 3ab7392..9c1b9af 100644
--- a/README.md
+++ b/README.md
@@ -31,6 +31,8 @@ in your IDE’s toolbar or build it directly from the terminal:
 
 ### Build and Run Desktop (JVM) Application
 
+Desktop offline English/Russian speech: [CPU TTS setup and audition](docs/desktop-tts.md).
+
 To build and run the development version of the desktop app, use the run configuration from the run widget
 in your IDE’s toolbar or run it directly from the terminal:
 
@@ -89,4 +91,4 @@ Learn more about [Kotlin Multiplatform](https://www.jetbrains.com/help/kotlin-mu
 
 We would appreciate your feedback on Compose/Web and Kotlin/Wasm in the public Slack
 channel [#compose-web](https://slack-chats.kotlinlang.org/c/compose-web).
-If you face any issues, please report them on [YouTrack](https://youtrack.jetbrains.com/newIssue?project=CMP).
\ No newline at end of file
+If you face any issues, please report them on [YouTrack](https://youtrack.jetbrains.com/newIssue?project=CMP).
diff --git a/composeApp/build.gradle.kts b/composeApp/build.gradle.kts
index 49fafcd..3ea49bf 100644
--- a/composeApp/build.gradle.kts
+++ b/composeApp/build.gradle.kts
@@ -96,10 +96,29 @@ kotlin {
         jvmMain.dependencies {
             implementation(compose.desktop.currentOs)
             implementation(libs.kotlinx.coroutinesSwing)
+            implementation(libs.onnxruntime)
+            implementation(libs.jackson.databind)
+        }
+        jvmTest.dependencies {
+            implementation(libs.kotlinx.coroutinesTest)
         }
     }
 }
 
+tasks.register<JavaExec>("ttsSample") {
+    group = "verification"
+    description = "Generate plain English/Russian WAV samples using the local Supertonic CPU model."
+    dependsOn("jvmMainClasses")
+    mainClass.set("net.sergeych.toread.tts.TtsSampleKt")
+    classpath = configurations["jvmRuntimeClasspath"] + kotlin.targets["jvm"].compilations["main"].output.allOutputs
+    workingDir = rootProject.projectDir
+    args(layout.buildDirectory.dir("tts-samples").get().asFile.absolutePath)
+    if (providers.gradleProperty("ttsPlay").orNull == "true") args("--play")
+    if (providers.gradleProperty("ttsStress").orNull == "true") args("--stress")
+    if (providers.gradleProperty("ttsRuAccent").orNull == "true") args("--ruaccent")
+    if (providers.gradleProperty("ttsSilero").orNull == "true") args("--silero")
+}
+
 android {
     namespace = "net.sergeych.toread"
     compileSdk = libs.versions.android.compileSdk.get().toInt()
diff --git a/composeApp/src/commonMain/kotlin/net/sergeych/toread/Localization.kt b/composeApp/src/commonMain/kotlin/net/sergeych/toread/Localization.kt
index 40241be..457f21c 100644
--- a/composeApp/src/commonMain/kotlin/net/sergeych/toread/Localization.kt
+++ b/composeApp/src/commonMain/kotlin/net/sergeych/toread/Localization.kt
@@ -174,6 +174,9 @@ internal open class AppStrings {
     open val offlineVoice = "Offline voice"
     open val autoRussian = "Auto Russian"
     open val loadingVoices = "Loading voices..."
+    open val preparingSpeech = "Preparing speech…"
+    open val offlineSpeechModelMissing = "Download the offline speech model first; see the desktop setup guide."
+    open val speechPlaybackFailed = "Speech playback failed. Check the offline model and audio output, then try again."
 
     open val imageCopied = "Image copied"
     open val copyFailed = "Copy failed"
@@ -399,6 +402,9 @@ internal object RussianStrings : AppStrings() {
     override val offlineVoice = "Офлайн-голос"
     override val autoRussian = "Авто: русский"
     override val loadingVoices = "Загружаем голоса..."
+    override val preparingSpeech = "Подготавливаем речь…"
+    override val offlineSpeechModelMissing = "Сначала загрузите офлайн-модель речи по инструкции для компьютера."
+    override val speechPlaybackFailed = "Не удалось воспроизвести речь. Проверьте офлайн-модель и вывод звука и попробуйте снова."
 
     override val imageCopied = "Изображение скопировано"
     override val copyFailed = "Не удалось скопировать"
diff --git a/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReadAloudPlatform.kt b/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReadAloudPlatform.kt
index 79e7c98..00f5dfb 100644
--- a/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReadAloudPlatform.kt
+++ b/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReadAloudPlatform.kt
@@ -17,6 +17,7 @@ data class ReadAloudState(
     val active: Boolean = false,
     val playing: Boolean = false,
     val sentenceIndex: Int? = null,
+    val buffering: Boolean = false,
 )
 
 data class ReadAloudEngineOption(
@@ -45,6 +46,7 @@ data class ReadAloudSettingsState(
     val textReplacements: List<ReadAloudTextReplacement> = emptyList(),
     val loading: Boolean = false,
     val message: String? = null,
+    val usesSystemDefaults: Boolean = true,
 )
 
 expect object ReadAloudPlatform {
diff --git a/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReaderScreen.kt b/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReaderScreen.kt
index 98c043e..ef8db06 100644
--- a/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReaderScreen.kt
+++ b/composeApp/src/commonMain/kotlin/net/sergeych/toread/ReaderScreen.kt
@@ -204,6 +204,10 @@ internal fun BookView(
         }
     }
 
+    LaunchedEffect(readAloudSettings.message) {
+        readAloudSettings.message?.let { snackbarHostState.showSnackbar(it) }
+    }
+
     fun openReaderLink(href: String) {
         if (href.isWebUrl()) {
             openPlatformUrl(href)
@@ -549,6 +553,7 @@ internal fun BookView(
                 if (readAloudPanelVisible && readAloudState.active) {
                     ReadAloudPanel(
                         playing = readAloudState.playing,
+                        buffering = readAloudState.buffering,
                         onPlayStop = {
                             if (readAloudState.playing) {
                                 ReadAloudPlatform.stop()
@@ -1626,6 +1631,7 @@ private val DarkReaderTextPalette = listOf(
 @Composable
 private fun ReadAloudPanel(
     playing: Boolean,
+    buffering: Boolean,
     onPlayStop: () -> Unit,
     onBack: () -> Unit,
     onForward: () -> Unit,
@@ -1642,6 +1648,10 @@ private fun ReadAloudPanel(
             horizontalArrangement = Arrangement.Center,
             verticalAlignment = Alignment.CenterVertically,
         ) {
+            if (buffering) {
+                CircularProgressIndicator(modifier = Modifier.size(20.dp))
+                Text(strings.preparingSpeech, modifier = Modifier.padding(horizontal = 8.dp))
+            }
             IconButton(onClick = onBack) {
                 Icon(Icons.Filled.Replay, contentDescription = strings.previousSentence)
             }
@@ -1686,7 +1696,7 @@ private fun ReadAloudSettingsDialog(
                         Text(selectedEngineLabel)
                     }
                     DropdownMenu(expanded = engineMenuOpen, onDismissRequest = { engineMenuOpen = false }) {
-                        DropdownMenuItem(
+                        if (state.usesSystemDefaults) DropdownMenuItem(
                             text = { Text(strings.systemDefault) },
                             onClick = {
                                 engineMenuOpen = false
@@ -1711,7 +1721,7 @@ private fun ReadAloudSettingsDialog(
                         Text(selectedVoiceLabel)
                     }
                     DropdownMenu(expanded = voiceMenuOpen, onDismissRequest = { voiceMenuOpen = false }) {
-                        DropdownMenuItem(
+                        if (state.usesSystemDefaults) DropdownMenuItem(
                             text = { Text(strings.autoRussian) },
                             onClick = {
                                 voiceMenuOpen = false
diff --git a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/ReadAloudPlatform.jvm.kt b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/ReadAloudPlatform.jvm.kt
index 6dba933..fa07cb3 100644
--- a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/ReadAloudPlatform.jvm.kt
+++ b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/ReadAloudPlatform.jvm.kt
@@ -1,20 +1,54 @@
 package net.sergeych.toread
 
-import kotlinx.coroutines.flow.MutableStateFlow
+import kotlinx.coroutines.CoroutineScope
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.SupervisorJob
+import kotlinx.coroutines.currentCoroutineContext
+import kotlinx.coroutines.ensureActive
 import kotlinx.coroutines.flow.StateFlow
+import kotlinx.coroutines.sync.Mutex
+import kotlinx.coroutines.sync.withLock
+import kotlinx.coroutines.withContext
+import net.sergeych.toread.tts.DesktopReadAloudPlayer
+import net.sergeych.toread.tts.SpeechAudio
+import net.sergeych.toread.tts.RuAccentProcess
+import net.sergeych.toread.tts.SileroProcess
+import net.sergeych.toread.tts.SupertonicTts
+import net.sergeych.toread.tts.playDesktopAudio
 
 actual object ReadAloudPlatform {
-    actual val isSupported: Boolean = false
-    private val mutableState = MutableStateFlow(ReadAloudState())
-    actual val state: StateFlow<ReadAloudState> = mutableState
-    private val mutableSettingsState = MutableStateFlow(ReadAloudSettingsState())
-    actual val settingsState: StateFlow<ReadAloudSettingsState> = mutableSettingsState
+    actual val isSupported: Boolean = true
+    private val inferenceLock = Mutex()
+    private var synthesizer: SupertonicTts? = null
+    private val accentor = if (RuAccentProcess.isInstalled()) RuAccentProcess() else null
+    private val silero = if (SileroProcess.isInstalled()) SileroProcess() else null
+    private val player = DesktopReadAloudPlayer(
+        scope = CoroutineScope(SupervisorJob() + Dispatchers.Main.immediate),
+        synthesize = { text, voice ->
+            withContext(Dispatchers.IO) {
+                inferenceLock.withLock {
+                    val context = currentCoroutineContext()
+                    context.ensureActive()
+                    val engine = synthesizer ?: SupertonicTts().also { synthesizer = it }
+                    context.ensureActive()
+                    SpeechAudio(engine.synthesize(text, voice, checkCancelled = { context.ensureActive() }), engine.sampleRate)
+                }
+            }
+        },
+        playAudio = ::playDesktopAudio,
+        installed = { SupertonicTts.isInstalled() },
+        accentuate = accentor?.let { processor -> { text -> processor.accentuate(text) } },
+        silero = silero?.let { processor -> { text, voice -> processor.synthesize(text, voice) } },
+    )
+    actual val state: StateFlow<ReadAloudState> = player.state
+    actual val settingsState: StateFlow<ReadAloudSettingsState> = player.settingsState
 
-    actual fun prepare(fileId: String, bookTitle: String, sentences: List<ReadAloudSentence>, startIndex: Int) = Unit
-    actual fun play() = Unit
-    actual fun stop() = Unit
-    actual fun skip(delta: Int) = Unit
-    actual fun refreshSettings() = Unit
-    actual fun selectEngine(engineId: String?) = Unit
-    actual fun selectVoice(voiceId: String?) = Unit
+    actual fun prepare(fileId: String, bookTitle: String, sentences: List<ReadAloudSentence>, startIndex: Int) =
+        player.prepare(sentences, startIndex)
+    actual fun play() = player.play()
+    actual fun stop() = player.stop()
+    actual fun skip(delta: Int) = player.skip(delta)
+    actual fun refreshSettings() = player.refreshSettings()
+    actual fun selectEngine(engineId: String?) = player.selectEngine(engineId)
+    actual fun selectVoice(voiceId: String?) = player.selectVoice(voiceId)
 }
diff --git a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/DesktopAudio.kt b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/DesktopAudio.kt
new file mode 100644
index 0000000..4875b31
--- /dev/null
+++ b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/DesktopAudio.kt
@@ -0,0 +1,70 @@
+package net.sergeych.toread.tts
+
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.currentCoroutineContext
+import kotlinx.coroutines.suspendCancellableCoroutine
+import kotlinx.coroutines.ensureActive
+import kotlinx.coroutines.withContext
+import java.io.ByteArrayInputStream
+import java.io.File
+import javax.sound.sampled.AudioFileFormat
+import javax.sound.sampled.AudioFormat
+import javax.sound.sampled.AudioInputStream
+import javax.sound.sampled.AudioSystem
+import javax.sound.sampled.LineEvent
+import java.util.concurrent.atomic.AtomicBoolean
+import kotlin.coroutines.resume
+
+internal fun SpeechAudio.pcm16(): ByteArray = ByteArray(samples.size * 2).also { bytes ->
+    samples.forEachIndexed { index, sample ->
+        val pcm = (sample.takeIf { it.isFinite() }?.coerceIn(-1f, 1f)?.times(32767) ?: 0f).toInt()
+        bytes[index * 2] = pcm.toByte()
+        bytes[index * 2 + 1] = (pcm shr 8).toByte()
+    }
+}
+
+internal suspend fun playDesktopAudio(audio: SpeechAudio) = withContext(Dispatchers.IO) {
+    val format = AudioFormat(audio.sampleRate.toFloat(), 16, 1, true, false)
+    val pcm = audio.pcm16()
+    val clip = AudioSystem.getClip()
+    try {
+        currentCoroutineContext().ensureActive()
+        clip.open(format, pcm, 0, pcm.size)
+        suspendCancellableCoroutine<Unit> { continuation ->
+            val completed = AtomicBoolean(false)
+            clip.addLineListener { event ->
+                if (event.type == LineEvent.Type.STOP && completed.compareAndSet(false, true)) {
+                    continuation.resume(Unit)
+                }
+            }
+            continuation.invokeOnCancellation {
+                clip.stop()
+                clip.close()
+            }
+            if (continuation.isActive) clip.start()
+        }
+    } finally {
+        clip.stop()
+        clip.close()
+    }
+}
+
+internal fun SpeechAudio.writeWav(file: File) {
+    file.parentFile?.mkdirs()
+    val format = AudioFormat(sampleRate.toFloat(), 16, 1, true, false)
+    AudioInputStream(ByteArrayInputStream(pcm16()), format, samples.size.toLong()).use {
+        AudioSystem.write(it, AudioFileFormat.Type.WAVE, file)
+    }
+}
+
+internal fun joinSpeechAudio(parts: List<SpeechAudio>): SpeechAudio {
+    require(parts.isNotEmpty()) { "No speech to play" }
+    if (parts.size == 1) return parts.single()
+    val rate = parts.first().sampleRate
+    require(parts.all { it.sampleRate == rate }) { "Mismatched speech sample rates" }
+    val pause = rate / 10
+    val samples = FloatArray(parts.sumOf { it.samples.size } + pause * (parts.size - 1))
+    var offset = 0
+    parts.forEach { part -> part.samples.copyInto(samples, offset); offset += part.samples.size + pause }
+    return SpeechAudio(samples, rate)
+}
diff --git a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/DesktopReadAloudPlayer.kt b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/DesktopReadAloudPlayer.kt
new file mode 100644
index 0000000..ce5f446
--- /dev/null
+++ b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/DesktopReadAloudPlayer.kt
@@ -0,0 +1,158 @@
+package net.sergeych.toread.tts
+
+import kotlinx.coroutines.CancellationException
+import kotlinx.coroutines.CoroutineScope
+import kotlinx.coroutines.Job
+import kotlinx.coroutines.async
+import kotlinx.coroutines.coroutineScope
+import kotlinx.coroutines.delay
+import kotlinx.coroutines.flow.MutableStateFlow
+import kotlinx.coroutines.flow.asStateFlow
+import kotlinx.coroutines.launch
+import net.sergeych.toread.ReadAloudEngineOption
+import net.sergeych.toread.ReadAloudSentence
+import net.sergeych.toread.ReadAloudSettingsState
+import net.sergeych.toread.ReadAloudState
+import net.sergeych.toread.ReadAloudVoiceOption
+import net.sergeych.toread.strings
+
+internal data class SpeechAudio(val samples: FloatArray, val sampleRate: Int)
+
+/** Commands and state changes run on the UI dispatcher; inference/audio run off-thread.
+ * A playback job owns its lookahead, so stop/skip/voice changes cannot publish stale audio.
+ */
+internal class DesktopReadAloudPlayer(
+    private val scope: CoroutineScope,
+    private val synthesize: suspend (String, String) -> SpeechAudio,
+    private val playAudio: suspend (SpeechAudio) -> Unit,
+    private val installed: () -> Boolean,
+    private val accentuate: (suspend (String) -> String)? = null,
+    private val silero: (suspend (String, String) -> SpeechAudio)? = null,
+) {
+    private fun voices(engine: String) = if (engine == "silero")
+        SileroProcess.voices.map { ReadAloudVoiceOption(it, it.replaceFirstChar(Char::uppercase), "ru (English: F1)", false) }
+    else SupertonicTts.voices.map { ReadAloudVoiceOption(it, it, "en / ru", false) }
+    private val defaultEngine = if (silero != null) "silero" else if (accentuate != null) "supertonic-3-ruaccent" else "supertonic-3"
+    private val mutableState = MutableStateFlow(ReadAloudState())
+    val state = mutableState.asStateFlow()
+    private val mutableSettings = MutableStateFlow(ReadAloudSettingsState(
+        engines = (if (silero != null) listOf(ReadAloudEngineOption("silero", "Silero (RU) / Supertonic (EN)")) else emptyList()) +
+            listOf(ReadAloudEngineOption("supertonic-3", "Supertonic 3")) +
+            if (accentuate != null) listOf(ReadAloudEngineOption("supertonic-3-ruaccent", "Supertonic 3 + RUAccent")) else emptyList(),
+        selectedEngineId = defaultEngine,
+        voices = voices(defaultEngine),
+        selectedVoiceId = if (silero != null) "xenia" else "F1",
+        usesSystemDefaults = false,
+    ))
+    val settingsState = mutableSettings.asStateFlow()
+    private var queue = emptyList<ReadAloudSentence>()
+    private var index = 0
+    private var playback: Job? = null
+
+    fun prepare(sentences: List<ReadAloudSentence>, startIndex: Int) {
+        stop()
+        queue = sentences.toList()
+        index = if (queue.isEmpty()) 0 else startIndex.coerceIn(queue.indices)
+        mutableState.value = ReadAloudState(active = queue.isNotEmpty(), sentenceIndex = queue.getOrNull(index)?.index)
+    }
+
+    fun play() {
+        if (queue.isEmpty()) return
+        playback?.cancel()
+        mutableSettings.value = mutableSettings.value.copy(message = null)
+        if (!installed()) {
+            mutableState.value = mutableState.value.copy(active = true, playing = false, buffering = false)
+            mutableSettings.value = mutableSettings.value.copy(message = strings.offlineSpeechModelMissing)
+            return
+        }
+        val voice = mutableSettings.value.selectedVoiceId ?: "F1"
+        val useSilero = mutableSettings.value.selectedEngineId == "silero"
+        val useAccents = mutableSettings.value.selectedEngineId == "supertonic-3-ruaccent"
+        suspend fun render(text: String): SpeechAudio {
+            if (useSilero) return if (SupertonicTts.languageOf(text) == "ru") silero!!.invoke(text, voice)
+                else synthesize(text, "F1")
+            val prepared = if (useAccents) accentuate!!.invoke(text) else text
+            // Accent marks increase character count; re-chunk before entering the TTS model.
+            return joinSpeechAudio(SupertonicTts.chunks(SupertonicTts.russianStressToUnicode(prepared)).map { synthesize(it, voice) })
+        }
+        val chunks = queue.asSequence().withIndex().drop(index).flatMap { (position, sentence) ->
+            val parts = SupertonicTts.chunks(sentence.spokenText)
+            parts.asSequence().mapIndexed { chunkIndex, text ->
+                Chunk(position, sentence, text, chunkIndex == 0, chunkIndex == parts.lastIndex)
+            }
+        }.iterator()
+        mutableState.value = mutableState.value.copy(active = true, playing = true, buffering = true)
+        playback = scope.launch {
+            try {
+                coroutineScope {
+                    // Keep at most one audio chunk ahead of the currently playing chunk.
+                    var next = if (chunks.hasNext()) chunks.next() else null
+                    var pending = next?.let { chunk -> async { render(chunk.text) } }
+                    while (next != null) {
+                        val current = next
+                        mutableState.value = mutableState.value.copy(buffering = pending?.isCompleted != true)
+                        val audio = pending!!.await()
+                        next = if (chunks.hasNext()) chunks.next() else null
+                        pending = next?.let { chunk -> async { render(chunk.text) } }
+                        index = current.position
+                        mutableState.value = ReadAloudState(active = true, playing = true, sentenceIndex = current.sentence.index)
+                        if (current.first) delay(current.sentence.pauseBeforeMillis)
+                        playAudio(audio)
+                        if (current.last) delay(current.sentence.pauseAfterMillis)
+                    }
+                }
+                mutableState.value = ReadAloudState()
+            } catch (cancelled: CancellationException) {
+                throw cancelled
+            } catch (error: Exception) {
+                System.err.println("Desktop TTS failed: ${error.javaClass.simpleName}: ${error.message}")
+                mutableState.value = mutableState.value.copy(playing = false, buffering = false)
+                mutableSettings.value = mutableSettings.value.copy(message = strings.speechPlaybackFailed)
+            }
+        }
+    }
+
+    fun stop() {
+        playback?.cancel()
+        playback = null
+        mutableState.value = ReadAloudState()
+    }
+
+    fun skip(delta: Int) {
+        if (queue.isEmpty()) return
+        val wasPlaying = mutableState.value.playing
+        playback?.cancel()
+        index = (index.toLong() + delta).coerceIn(0, queue.lastIndex.toLong()).toInt()
+        mutableState.value = ReadAloudState(active = true, sentenceIndex = queue[index].index)
+        if (wasPlaying) play()
+    }
+
+    fun selectVoice(id: String?) {
+        val voice = id ?: if (mutableSettings.value.selectedEngineId == "silero") "xenia" else "F1"
+        if (mutableSettings.value.voices.none { it.id == voice } || voice == mutableSettings.value.selectedVoiceId) return
+        mutableSettings.value = mutableSettings.value.copy(selectedVoiceId = voice)
+        if (mutableState.value.playing) play()
+    }
+
+    fun selectEngine(id: String?) {
+        val engine = id ?: defaultEngine
+        if (mutableSettings.value.engines.none { it.id == engine } || engine == mutableSettings.value.selectedEngineId) return
+        val options = voices(engine)
+        mutableSettings.value = mutableSettings.value.copy(selectedEngineId = engine, voices = options,
+            selectedVoiceId = mutableSettings.value.selectedVoiceId?.takeIf { id -> options.any { it.id == id } }
+                ?: if (engine == "silero") "xenia" else "F1")
+        if (mutableState.value.playing) play()
+    }
+
+    fun refreshSettings() {
+        mutableSettings.value = mutableSettings.value.copy(message = if (installed()) null else strings.offlineSpeechModelMissing)
+    }
+
+    private data class Chunk(
+        val position: Int,
+        val sentence: ReadAloudSentence,
+        val text: String,
+        val first: Boolean,
+        val last: Boolean,
+    )
+}
diff --git a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/RuAccentProcess.kt b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/RuAccentProcess.kt
new file mode 100644
index 0000000..dee0afa
--- /dev/null
+++ b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/RuAccentProcess.kt
@@ -0,0 +1,90 @@
+package net.sergeych.toread.tts
+
+import com.fasterxml.jackson.databind.ObjectMapper
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.async
+import kotlinx.coroutines.coroutineScope
+import kotlinx.coroutines.currentCoroutineContext
+import kotlinx.coroutines.ensureActive
+import kotlinx.coroutines.isActive
+import kotlinx.coroutines.sync.Mutex
+import kotlinx.coroutines.sync.withLock
+import kotlinx.coroutines.withTimeout
+import java.io.BufferedReader
+import java.io.BufferedWriter
+import java.io.File
+
+/** Optional desktop-only CPU accentor. Persistent local process, no HTTP or remote text service. */
+internal class RuAccentProcess(private val directory: File = installationDirectory()) : AutoCloseable {
+    private val mapper = ObjectMapper()
+    private val lock = Mutex()
+    @Volatile private var worker: Worker? = null
+
+    init {
+        Runtime.getRuntime().addShutdownHook(Thread { close() })
+    }
+
+    suspend fun accentuate(text: String): String {
+        // Respect explicitly supplied stress; English needs no Russian processor.
+        if (SupertonicTts.languageOf(text) == "en" || '\u0301' in text || Regex("\\+[аеёиоуыэюяАЕЁИОУЫЭЮЯ]").containsMatchIn(text)) return text
+        return lock.withLock {
+            try {
+                currentCoroutineContext().ensureActive()
+                val current = worker?.takeIf { it.process.isAlive } ?: startWorker().also { worker = it }
+                withTimeout(90_000) {
+                    coroutineScope {
+                        val response = async(Dispatchers.IO) {
+                            if (!current.ready) {
+                                val ready = current.output.readLine() ?: error("RuAccent exited while loading")
+                                check(mapper.readTree(ready)["ready"]?.asBoolean() == true) { "RuAccent did not initialize" }
+                                current.ready = true
+                            }
+                            current.input.write(mapper.writeValueAsString(mapOf("text" to text)))
+                            current.input.newLine()
+                            current.input.flush()
+                            val line = current.output.readLine() ?: error("RuAccent exited during inference")
+                            val result = mapper.readTree(line)
+                            check(result["error"] == null) { "RuAccent inference failed" }
+                            result["text"]?.asText()?.takeIf { it.isNotBlank() } ?: error("Empty RuAccent response")
+                        }
+                        try {
+                            response.await()
+                        } finally {
+                            // A blocked pipe read cannot be interrupted reliably; killing the
+                            // cancelled worker unblocks it before the next request acquires the lock.
+                            if (!isActive) current.process.destroyForcibly()
+                        }
+                    }
+                }
+            } catch (error: Exception) {
+                close()
+                throw error
+            }
+        }
+    }
+
+    private fun startWorker(): Worker {
+        val runtime = mapper.readTree(File(directory, "runtime.json"))
+        val process = ProcessBuilder(runtime["python"].asText(), "-u", File(directory, "worker.py").path, directory.path)
+            .redirectError(ProcessBuilder.Redirect.INHERIT)
+            .apply {
+                environment()["HF_HUB_OFFLINE"] = "1"
+                environment()["TRANSFORMERS_OFFLINE"] = "1"
+                environment()["HF_HUB_DISABLE_TELEMETRY"] = "1"
+            }.start()
+        return Worker(process, process.inputStream.bufferedReader(Charsets.UTF_8), process.outputStream.bufferedWriter(Charsets.UTF_8))
+    }
+
+    override fun close() {
+        worker?.process?.destroyForcibly()
+        worker = null
+    }
+
+    private class Worker(val process: Process, val output: BufferedReader, val input: BufferedWriter, var ready: Boolean = false)
+
+    companion object {
+        fun installationDirectory(): File = System.getenv("YEREED_RUACCENT_HOME")?.let(::File)
+            ?: File(SupertonicTts.modelDirectory().parentFile, "ruaccent")
+        fun isInstalled(): Boolean = File(installationDirectory(), "runtime.json").isFile && File(installationDirectory(), "worker.py").isFile
+    }
+}
diff --git a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/SileroProcess.kt b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/SileroProcess.kt
new file mode 100644
index 0000000..94e8cfb
--- /dev/null
+++ b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/SileroProcess.kt
@@ -0,0 +1,98 @@
+package net.sergeych.toread.tts
+
+import com.fasterxml.jackson.databind.ObjectMapper
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.async
+import kotlinx.coroutines.coroutineScope
+import kotlinx.coroutines.currentCoroutineContext
+import kotlinx.coroutines.ensureActive
+import kotlinx.coroutines.isActive
+import kotlinx.coroutines.sync.Mutex
+import kotlinx.coroutines.sync.withLock
+import kotlinx.coroutines.withTimeout
+import java.io.BufferedReader
+import java.io.BufferedWriter
+import java.io.File
+import java.nio.ByteBuffer
+import java.nio.ByteOrder
+import java.util.Base64
+
+/** Desktop-only CPU Silero synthesizer. Persistent local process, no HTTP or remote text service. */
+internal class SileroProcess(private val directory: File = installationDirectory()) : AutoCloseable {
+    private val mapper = ObjectMapper()
+    private val lock = Mutex()
+    @Volatile private var worker: Worker? = null
+
+    init {
+        Runtime.getRuntime().addShutdownHook(Thread { close() })
+    }
+
+    suspend fun synthesize(text: String, voice: String): SpeechAudio {
+        return lock.withLock {
+            try {
+                currentCoroutineContext().ensureActive()
+                val current = worker?.takeIf { it.process.isAlive } ?: startWorker().also { worker = it }
+                withTimeout(90_000) {
+                    coroutineScope {
+                        val response = async(Dispatchers.IO) {
+                            if (!current.ready) {
+                                val ready = current.output.readLine() ?: error("Silero exited while loading")
+                                check(mapper.readTree(ready)["ready"]?.asBoolean() == true) { "Silero did not initialize" }
+                                current.ready = true
+                            }
+                            current.input.write(mapper.writeValueAsString(mapOf("text" to text, "voice" to voice)))
+                            current.input.newLine()
+                            current.input.flush()
+                            val line = current.output.readLine() ?: error("Silero exited during inference")
+                            val result = mapper.readTree(line)
+                            check(result["error"] == null) { "Silero inference failed" }
+                            val bytes = Base64.getDecoder().decode(result["audio"].asText())
+                            check(bytes.isNotEmpty() && bytes.size % 4 == 0) { "Invalid Silero audio" }
+                            val buffer = ByteBuffer.wrap(bytes).order(ByteOrder.LITTLE_ENDIAN).asFloatBuffer()
+                            val samples = FloatArray(buffer.remaining()).also { buffer.get(it) }
+                            check(samples.all { it.isFinite() }) { "Non-finite Silero audio" }
+                            SpeechAudio(samples, result["sampleRate"].asInt().also { check(it == 48000) })
+                        }
+                        try {
+                            response.await()
+                        } finally {
+                            // A blocked pipe read cannot be interrupted reliably; killing the
+                            // cancelled worker unblocks it before the next request acquires the lock.
+                            if (!isActive) current.process.destroyForcibly()
+                        }
+                    }
+                }
+            } catch (error: Exception) {
+                close()
+                throw error
+            }
+        }
+    }
+
+    private fun startWorker(): Worker {
+        val runtime = mapper.readTree(File(directory, "runtime.json"))
+        val script = SileroProcess::class.java.getResourceAsStream("/tts/silero-worker.py")!!.use { it.reader().readText() }
+        val process = ProcessBuilder(runtime["python"].asText(), "-u", "-c", script, directory.path)
+            .redirectError(ProcessBuilder.Redirect.INHERIT)
+            .apply {
+                environment()["HF_HUB_OFFLINE"] = "1"
+                environment()["TRANSFORMERS_OFFLINE"] = "1"
+                environment()["HF_HUB_DISABLE_TELEMETRY"] = "1"
+            }.start()
+        return Worker(process, process.inputStream.bufferedReader(Charsets.UTF_8), process.outputStream.bufferedWriter(Charsets.UTF_8))
+    }
+
+    override fun close() {
+        worker?.process?.destroyForcibly()
+        worker = null
+    }
+
+    private class Worker(val process: Process, val output: BufferedReader, val input: BufferedWriter, var ready: Boolean = false)
+
+    companion object {
+        val voices = listOf("xenia", "baya", "aidar", "kseniya", "eugene")
+        fun installationDirectory(): File = System.getenv("YEREED_SILERO_HOME")?.let(::File)
+            ?: File(SupertonicTts.modelDirectory().parentFile, "silero-v5_5-ru")
+        fun isInstalled(): Boolean = File(installationDirectory(), "runtime.json").isFile && File(installationDirectory(), "model.pt").length() > 0
+    }
+}
diff --git a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/SupertonicTts.kt b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/SupertonicTts.kt
new file mode 100644
index 0000000..229beba
--- /dev/null
+++ b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/SupertonicTts.kt
@@ -0,0 +1,190 @@
+package net.sergeych.toread.tts
+
+import ai.onnxruntime.OnnxTensor
+import ai.onnxruntime.OrtEnvironment
+import ai.onnxruntime.OrtSession
+import com.fasterxml.jackson.databind.JsonNode
+import com.fasterxml.jackson.databind.ObjectMapper
+import java.io.File
+import java.nio.FloatBuffer
+import java.text.Normalizer
+import java.util.Random
+
+/** CPU-only, single-utterance adaptation of Supertone's MIT Java example.
+ * See docs/licenses/supertonic-code-MIT.txt. Call from one worker at a time.
+ */
+internal class SupertonicTts(val directory: File = modelDirectory()) : AutoCloseable {
+    private val mapper = ObjectMapper()
+    private val environment = OrtEnvironment.getEnvironment()
+    private val config = mapper.readTree(File(directory, "onnx/tts.json"))
+    private val indexer = mapper.readTree(File(directory, "onnx/unicode_indexer.json"))
+        .map { it.asLong() }.toLongArray()
+    val sampleRate: Int = config["ae"]["sample_rate"].asInt()
+    private val compression = config["ttl"]["chunk_compress_factor"].asInt()
+    private val chunkSize = config["ae"]["base_chunk_size"].asInt() * compression
+    private val latentDimension = config["ttl"]["latent_dim"].asInt() * compression
+    private val sessions = mutableListOf<OrtSession>()
+    private val styles = mutableMapOf<String, JsonNode>()
+
+    init {
+        try {
+            OrtSession.SessionOptions().use { options ->
+                // The standard (non-GPU) ONNX artifact uses CPUExecutionProvider only.
+                options.setIntraOpNumThreads(4.coerceAtMost(Runtime.getRuntime().availableProcessors()))
+                options.setInterOpNumThreads(1)
+                for (name in listOf("duration_predictor", "text_encoder", "vector_estimator", "vocoder")) {
+                    sessions += environment.createSession(File(directory, "onnx/$name.onnx").path, options)
+                }
+            }
+        } catch (error: Throwable) {
+            close()
+            throw error
+        }
+    }
+
+    fun synthesize(
+        text: String,
+        voice: String = "F1",
+        language: String = languageOf(text),
+        seed: Long? = null,
+        checkCancelled: () -> Unit = {},
+    ): FloatArray {
+        require(voice in voices) { "Unknown voice: $voice" }
+        require(language in listOf("en", "ru", "na")) { "Unsupported language: $language" }
+        require(text.isNotBlank() && text.length <= MAX_CHARS) { "Expected 1–$MAX_CHARS characters" }
+        checkCancelled()
+        val prepared = prepareText(text, language)
+        // Unknown symbols become spaces, never invalid (-1) embedding indices.
+        val ids = prepared.codePoints().toArray().map { point ->
+            indexer.getOrNull(point)?.takeIf { it >= 0 } ?: indexer[' '.code]
+        }.toLongArray()
+        val style = styles.getOrPut(voice) { mapper.readTree(File(directory, "voice_styles/$voice.json")) }
+        NativeResources().use { resources ->
+            fun tensor(value: Any) = resources.keep(OnnxTensor.createTensor(environment, value))
+            fun styleTensor(name: String): OnnxTensor {
+                val node = style[name]
+                val values = node["data"].flatMap { batch -> batch.flatMap { row -> row.map { it.floatValue() } } }
+                return resources.keep(OnnxTensor.createTensor(
+                    environment, FloatBuffer.wrap(values.toFloatArray()), node["dims"].map { it.asLong() }.toLongArray(),
+                ))
+            }
+            val textIds = tensor(arrayOf(ids))
+            val textMask = tensor(arrayOf(arrayOf(FloatArray(ids.size) { 1f })))
+            val styleDp = styleTensor("style_dp")
+            val styleTtl = styleTensor("style_ttl")
+            val duration = sessions[0].run(mapOf(
+                "text_ids" to textIds, "text_mask" to textMask, "style_dp" to styleDp,
+            )).use { result -> (result[0] as OnnxTensor).floatBuffer.get(0) / 1.05f }
+            check(duration.isFinite() && duration > 0 && duration <= 60) { "Invalid speech duration: $duration" }
+            checkCancelled()
+            val sampleCount = (duration * sampleRate).toInt()
+            val latentLength = (sampleCount + chunkSize - 1) / chunkSize
+            val random = seed?.let(::Random) ?: Random()
+            var latent = arrayOf(Array(latentDimension) { FloatArray(latentLength) { random.nextGaussian().toFloat() } })
+            val latentMask = tensor(arrayOf(arrayOf(FloatArray(latentLength) { 1f })))
+            val totalSteps = tensor(floatArrayOf(8f))
+            val encoded = resources.keep(sessions[1].run(mapOf(
+                "text_ids" to textIds, "text_mask" to textMask, "style_ttl" to styleTtl,
+            )))
+            val embedding = encoded[0] as OnnxTensor
+            repeat(8) { step ->
+                checkCancelled()
+                OnnxTensor.createTensor(environment, latent).use { noisy ->
+                    OnnxTensor.createTensor(environment, floatArrayOf(step.toFloat())).use { currentStep ->
+                        sessions[2].run(mapOf(
+                            "noisy_latent" to noisy, "text_emb" to embedding, "style_ttl" to styleTtl,
+                            "latent_mask" to latentMask, "text_mask" to textMask,
+                            "current_step" to currentStep, "total_step" to totalSteps,
+                        )).use { result ->
+                            @Suppress("UNCHECKED_CAST")
+                            latent = result[0].value as Array<Array<FloatArray>>
+                        }
+                    }
+                }
+            }
+            checkCancelled()
+            OnnxTensor.createTensor(environment, latent).use { finalLatent ->
+                return sessions[3].run(mapOf("latent" to finalLatent)).use { result ->
+                    val samples = (result[0] as OnnxTensor).floatBuffer
+                    check(samples.remaining() >= sampleCount) { "Truncated speech waveform" }
+                    FloatArray(sampleCount).also { samples.get(it) }
+                }
+            }
+        }
+    }
+
+    override fun close() {
+        sessions.asReversed().forEach { runCatching { it.close() } }
+        sessions.clear()
+    }
+
+    companion object {
+        const val MAX_CHARS = 300
+        val voices = listOf("F1", "F2", "F3", "F4", "F5", "M1", "M2", "M3", "M4", "M5")
+
+        fun modelDirectory(): File {
+            val override = System.getProperty("yereed.tts.modelDir") ?: System.getenv("YEREED_TTS_MODEL_DIR")
+            if (!override.isNullOrBlank()) return File(override)
+            generateSequence(File(System.getProperty("user.dir")).absoluteFile) { it.parentFile }.forEach { root ->
+                File(root, ".local/tts/supertonic-3").takeIf { it.isDirectory }?.let { return it }
+            }
+            return File(System.getProperty("user.home"), ".toread/tts/supertonic-3")
+        }
+
+        fun isInstalled(directory: File = modelDirectory()): Boolean =
+            (listOf("onnx/tts.json", "onnx/unicode_indexer.json") +
+                listOf("duration_predictor", "text_encoder", "vector_estimator", "vocoder").map { "onnx/$it.onnx" } +
+                voices.map { "voice_styles/$it.json" }).all { File(directory, it).length() > 0 }
+
+        fun languageOf(text: String): String {
+            val russian = text.any { it in 'А'..'я' || it == 'ё' || it == 'Ё' }
+            val english = text.any { it in 'a'..'z' || it in 'A'..'Z' }
+            return when {
+                russian && english -> "na"
+                russian -> "ru"
+                else -> "en"
+            }
+        }
+
+        /** Bounded chunks preserve every non-whitespace character, including long unbroken words. */
+        fun chunks(text: String): List<String> = buildList {
+            var remaining = text.trim()
+            while (remaining.length > MAX_CHARS) {
+                var end = remaining.lastIndexOfAny(charArrayOf(' ', '\n', '\t'), MAX_CHARS)
+                if (end <= 0) end = MAX_CHARS
+                if (Character.isHighSurrogate(remaining[end - 1])) end--
+                while (end > 0 && Character.getType(remaining[end]) == Character.NON_SPACING_MARK.toInt()) end--
+                require(end > 0) { "Oversized combining-character sequence" }
+                add(remaining.substring(0, end).trim())
+                remaining = remaining.substring(end).trimStart()
+            }
+            if (remaining.isNotEmpty()) add(remaining)
+        }
+
+        internal fun prepareText(text: String, language: String): String {
+            var normalized = Normalizer.normalize(russianStressToUnicode(text), Normalizer.Form.NFKD)
+                .replace(Regex("[–‑—]"), "-")
+                .replace(Regex("[“”]"), "\"")
+                .replace(Regex("[‘’´`]"), "'")
+                .replace(Regex("[_\\[\\]|/#→←<>]"), " ")
+                .replace(Regex("[♥☆♡©\\\\]"), "")
+                .replace(Regex("[\\x{1F000}-\\x{1FAFF}\\x{2600}-\\x{27BF}]"), "")
+            if (language == "en") normalized = normalized.replace("@", " at ")
+                .replace("e.g.,", "for example, ").replace("i.e.,", "that is, ")
+            normalized = normalized.replace(Regex("\\s+"), " ").trim()
+                .replace(Regex(" +([,.!?;:'])"), "$1")
+            if (normalized.isEmpty() || normalized.last() !in ".!?;:,'\")]}…»") normalized += "."
+            return "<$language>$normalized</$language>"
+        }
+
+        /** RUAccent/Silero notation -> Cyrillic vowel plus U+0301, never a Latin lookalike. */
+        internal fun russianStressToUnicode(text: String): String =
+            text.replace(Regex("\\+([аеёиоуыэюяАЕЁИОУЫЭЮЯ])")) { it.groupValues[1] + "\u0301" }
+    }
+}
+
+private class NativeResources : AutoCloseable {
+    private val values = mutableListOf<AutoCloseable>()
+    fun <T : AutoCloseable> keep(value: T): T = value.also { values += it }
+    override fun close() = values.asReversed().forEach { runCatching { it.close() } }
+}
diff --git a/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/TtsSample.kt b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/TtsSample.kt
new file mode 100644
index 0000000..2edbbb1
--- /dev/null
+++ b/composeApp/src/jvmMain/kotlin/net/sergeych/toread/tts/TtsSample.kt
@@ -0,0 +1,74 @@
+package net.sergeych.toread.tts
+
+import kotlinx.coroutines.runBlocking
+import java.io.File
+import java.util.Locale
+import kotlin.system.measureNanoTime
+
+/** Reproducible audition without opening the library, altering books, or contacting a service. */
+fun main(args: Array<String>) = runBlocking {
+    val output = File(args.firstOrNull() ?: "build/tts-samples")
+    if ("--silero" in args) {
+        check(SileroProcess.isInstalled()) { "Install the Silero desktop runtime first." }
+        SileroProcess().use { engine ->
+            for ((label, text) in listOf(
+                "library" to "На двери висел старый замок, а вдали виднелся замок на холме.",
+                "yo" to "Ежик шел по лесу. Ребенок пьет теплое молоко.",
+            )) {
+                val audio = engine.synthesize(text, "xenia")
+                check(audio.samples.any { kotlin.math.abs(it) > 0.001f })
+                val file = File(output, "silero-desktop-$label.wav")
+                audio.writeWav(file)
+                println("Silero desktop worker: ${file.absolutePath}; ${audio.samples.size.toDouble() / audio.sampleRate}s")
+                if ("--play" in args) playDesktopAudio(audio)
+            }
+        }
+        return@runBlocking
+    }
+    check(SupertonicTts.isInstalled()) { "Run python3 scripts/download-tts-model.py first." }
+    lateinit var engine: SupertonicTts
+    val loadTime = measureNanoTime { engine = SupertonicTts() } / 1e9
+    println("Supertonic 3; CPU only; up to 4 threads; 8 steps; model load: %.3f s".format(Locale.ROOT, loadTime))
+    val stressTrial = "--stress" in args
+    val ruaccentTrial = "--ruaccent" in args
+    val russianText = "В библиотеке было тихо. Она открыла книгу и начала читать. За окном шёл осенний дождь. На двери висел старый замок, а вдали виднелся замок на холме. Завтра, подумала она, мы наконец доберёмся до моря."
+    val accentor = if (ruaccentTrial) RuAccentProcess() else null
+    val samples = if (ruaccentTrial) linkedMapOf("ru-plain" to russianText, "ru-ruaccent" to russianText) else if (stressTrial) linkedMapOf(
+        "ru-plain" to "Перед нами старый замок. Они стояли около замка.",
+        "ru-castle" to "П+еред н+ами ст+арый з+амок. Он+и сто+яли +около з+амка.",
+        "ru-lock" to "П+еред н+ами ст+арый зам+ок. Он+и сто+яли +около замк+а.",
+    ) else linkedMapOf(
+        "en" to "The library was quiet. She opened the book and began to read. Outside, the autumn rain tapped softly against the window. Tomorrow, she thought, they would finally reach the coast.",
+        "ru" to russianText,
+    )
+    engine.use {
+        try {
+            for (voice in if (stressTrial || ruaccentTrial) listOf("F1") else listOf("F1", "M1")) for ((label, text) in samples) {
+                val language = if (stressTrial || ruaccentTrial) "ru" else label
+                val accentStart = System.nanoTime()
+                val accented = if (label == "ru-ruaccent") accentor!!.accentuate(text) else text
+                if (label == "ru-ruaccent") println("RuAccent (including first load): %.3f s".format(Locale.ROOT, (System.nanoTime() - accentStart) / 1e9))
+                val spoken = SupertonicTts.russianStressToUnicode(accented)
+                lateinit var audio: SpeechAudio
+                val elapsed = measureNanoTime {
+                    audio = joinSpeechAudio(SupertonicTts.chunks(spoken).map { chunk ->
+                        SpeechAudio(engine.synthesize(chunk, voice, language, seed = 42), engine.sampleRate)
+                    })
+                } / 1e9
+                val duration = audio.samples.size.toDouble() / audio.sampleRate
+                check(audio.samples.all { it.isFinite() } && audio.samples.any { kotlin.math.abs(it) > 0.001f }) {
+                    "Invalid or silent model output"
+                }
+                val file = File(output, "$label-$voice.wav")
+                audio.writeWav(file)
+                File(output, "$label-$voice.txt").writeText(spoken + "\n")
+                println("$label/$voice: %.3f s synthesis, %.3f s audio, RTF %.3f; %s".format(
+                    Locale.ROOT, elapsed, duration, elapsed / duration, file.absolutePath,
+                ))
+                if ("--play" in args) playDesktopAudio(audio)
+            }
+        } finally {
+            accentor?.close()
+        }
+    }
+}
diff --git a/composeApp/src/jvmMain/resources/META-INF/licenses/supertonic-code-MIT.txt b/composeApp/src/jvmMain/resources/META-INF/licenses/supertonic-code-MIT.txt
new file mode 100644
index 0000000..943171b
--- /dev/null
+++ b/composeApp/src/jvmMain/resources/META-INF/licenses/supertonic-code-MIT.txt
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Supertone Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
\ No newline at end of file
diff --git a/composeApp/src/jvmMain/resources/tts/ruaccent-worker.py b/composeApp/src/jvmMain/resources/tts/ruaccent-worker.py
new file mode 100644
index 0000000..daaf700
--- /dev/null
+++ b/composeApp/src/jvmMain/resources/tts/ruaccent-worker.py
@@ -0,0 +1,56 @@
+"""Local stdin/stdout JSON protocol. Book text never leaves this process tree."""
+import contextlib
+import json
+import os
+from pathlib import Path
+import sys
+import time
+
+os.environ.update(HF_HUB_OFFLINE="1", TRANSFORMERS_OFFLINE="1", HF_HUB_DISABLE_TELEMETRY="1", TOKENIZERS_PARALLELISM="false")
+
+
+def load(model):
+    # Bound ORT worker counts before RUAccent imports its session constructor.
+    import onnxruntime as ort
+    original = ort.InferenceSession
+
+    def cpu_session(path, *args, **kwargs):
+        options = ort.SessionOptions()
+        options.intra_op_num_threads = 4
+        options.inter_op_num_threads = 1
+        kwargs["sess_options"] = options
+        kwargs["providers"] = ["CPUExecutionProvider"]
+        return original(path, *args, **kwargs)
+
+    ort.InferenceSession = cpu_session
+    from ruaccent import RUAccent
+    accentor = RUAccent()
+    accentor.load(omograph_model_size="turbo3.1", use_dictionary=True, device="CPU", workdir=str(model), tiny_mode=False)
+    return accentor
+
+
+def main():
+    with contextlib.redirect_stdout(sys.stderr):
+        accentor = load(Path(sys.argv[1]))
+    if "--check" in sys.argv:
+        with contextlib.redirect_stdout(sys.stderr):
+            result = accentor.process_all("На двери висит замок.")
+        print(result, flush=True)
+        return
+    print(json.dumps({"ready": True}), flush=True)
+    for line in sys.stdin:
+        try:
+            text = json.loads(line)["text"]
+            if not isinstance(text, str) or len(text) > 2000:
+                raise ValueError("Expected text of at most 2000 characters")
+            start = time.monotonic()
+            with contextlib.redirect_stdout(sys.stderr):
+                accented = accentor.process_all(text)
+            print(json.dumps({"text": accented, "seconds": time.monotonic() - start}, ensure_ascii=False), flush=True)
+        except Exception as error:
+            # Do not echo private input or exception text into application logs.
+            print(json.dumps({"error": type(error).__name__}), flush=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/composeApp/src/jvmMain/resources/tts/silero-worker.py b/composeApp/src/jvmMain/resources/tts/silero-worker.py
new file mode 100644
index 0000000..3f477c1
--- /dev/null
+++ b/composeApp/src/jvmMain/resources/tts/silero-worker.py
@@ -0,0 +1,47 @@
+"""Persistent CPU-only Silero worker; local JSON lines, float32 audio in base64."""
+import array
+import base64
+import contextlib
+import json
+import os
+from pathlib import Path
+import re
+import sys
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+os.environ["TORCH_HOME"] = str(Path(sys.argv[1]).parent / "torch")
+
+
+def main():
+    with contextlib.redirect_stdout(sys.stderr):
+        import torch
+        torch.set_num_threads(4)
+        torch.set_num_interop_threads(1)
+        model = torch.package.PackageImporter(str(Path(sys.argv[1]) / "model.pt")).load_pickle("tts_models", "model")
+        model.to(torch.device("cpu"))
+    print(json.dumps({"ready": True}), flush=True)
+    for line in sys.stdin:
+        try:
+            request = json.loads(line)
+            text = request["text"]
+            voice = request["voice"]
+            if not isinstance(text, str) or not 0 < len(text) <= 2000 or voice not in model.speakers:
+                raise ValueError("Invalid speech request")
+            # Preserve explicit reader/book stress in Silero's native notation.
+            text = re.sub(r"([аеёиоуыэюяАЕЁИОУЫЭЮЯ])\u0301", r"+\1", text)
+            with contextlib.redirect_stdout(sys.stderr), torch.inference_mode():
+                audio = model.apply_tts(text=text, speaker=voice, sample_rate=48000,
+                                        put_accent=True, put_yo=True,
+                                        put_stress_homo=True, put_yo_homo=True).detach().cpu().flatten()
+            if not audio.numel() or not torch.isfinite(audio).all():
+                raise ValueError("Invalid model audio")
+            samples = array.array("f", audio.tolist())
+            if sys.byteorder != "little":
+                samples.byteswap()
+            print(json.dumps({"sampleRate": 48000, "audio": base64.b64encode(samples.tobytes()).decode("ascii")}), flush=True)
+        except Exception as error:
+            print(json.dumps({"error": type(error).__name__}), flush=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/composeApp/src/jvmTest/kotlin/net/sergeych/toread/tts/DesktopReadAloudPlayerTest.kt b/composeApp/src/jvmTest/kotlin/net/sergeych/toread/tts/DesktopReadAloudPlayerTest.kt
new file mode 100644
index 0000000..8f3fec7
--- /dev/null
+++ b/composeApp/src/jvmTest/kotlin/net/sergeych/toread/tts/DesktopReadAloudPlayerTest.kt
@@ -0,0 +1,186 @@
+package net.sergeych.toread.tts
+
+import kotlinx.coroutines.ExperimentalCoroutinesApi
+import kotlinx.coroutines.awaitCancellation
+import kotlinx.coroutines.delay
+import kotlinx.coroutines.test.advanceUntilIdle
+import kotlinx.coroutines.test.runCurrent
+import kotlinx.coroutines.test.runTest
+import net.sergeych.toread.ReadAloudSentence
+import kotlin.test.Test
+import kotlin.test.assertEquals
+import kotlin.test.assertFalse
+import kotlin.test.assertNotNull
+import kotlin.test.assertNull
+import kotlin.test.assertTrue
+
+@OptIn(ExperimentalCoroutinesApi::class)
+class DesktopReadAloudPlayerTest {
+    private fun sentence(index: Int, text: String) = ReadAloudSentence(index, index, 0, text.length, text)
+    private fun audio(id: Int) = SpeechAudio(floatArrayOf(id.toFloat()), 24000)
+
+    @Test
+    fun sileroDefaultsToXeniaRoutesEnglishAndResetsVoiceOnEngineChange() = runTest {
+        val calls = mutableListOf<String>()
+        var accented = false
+        val player = DesktopReadAloudPlayer(this, { text, voice -> calls += "super:$voice:$text"; audio(1) },
+            {}, { true }, { accented = true; it }, { text, voice -> calls += "silero:$voice:$text"; audio(1) })
+        assertEquals("silero", player.settingsState.value.selectedEngineId)
+        assertEquals("xenia", player.settingsState.value.selectedVoiceId)
+        player.prepare(listOf(sentence(0, "Привет."), sentence(1, "Hello."), sentence(2, "Привет, John.")), 0)
+        player.play()
+        advanceUntilIdle()
+        assertEquals(listOf("silero:xenia:Привет.", "super:F1:Hello.", "super:F1:Привет, John."), calls)
+        assertFalse(accented)
+        player.selectVoice("F1")
+        assertEquals("xenia", player.settingsState.value.selectedVoiceId)
+        player.selectEngine("supertonic-3")
+        assertEquals("F1", player.settingsState.value.selectedVoiceId)
+        player.selectEngine("silero")
+        assertEquals("xenia", player.settingsState.value.selectedVoiceId)
+    }
+
+    @Test
+    fun switchingAwayFromSileroCancelsPendingSpeech() = runTest {
+        var cancelled = false
+        var played = 0
+        val player = DesktopReadAloudPlayer(this, { _, _ -> audio(1) }, { played++ }, { true },
+            silero = { _, _ -> try { awaitCancellation() } finally { cancelled = true } })
+        player.prepare(listOf(sentence(0, "Привет.")), 0)
+        player.play()
+        runCurrent()
+        player.selectEngine("supertonic-3")
+        advanceUntilIdle()
+        assertTrue(cancelled)
+        assertEquals(1, played)
+    }
+
+    @Test
+    fun stopDuringSynthesisDiscardsAudioAndDoesNotReportAnError() = runTest {
+        var started = false
+        var played = false
+        val player = DesktopReadAloudPlayer(this, { _, _ -> started = true; awaitCancellation() }, { played = true }, { true })
+        player.prepare(listOf(sentence(7, "First.")), 0)
+        player.play()
+        runCurrent()
+        assertTrue(started)
+        assertTrue(player.state.value.buffering)
+        player.stop()
+        runCurrent()
+        assertFalse(played)
+        assertFalse(player.state.value.active)
+        assertNull(player.settingsState.value.message)
+    }
+
+    @Test
+    fun skipCancelsOldSynthesisAndStartsRequestedSentence() = runTest {
+        val played = mutableListOf<Int>()
+        val player = DesktopReadAloudPlayer(this, { text, _ ->
+            if (text == "First.") awaitCancellation() else audio(2)
+        }, { played += it.samples[0].toInt(); awaitCancellation() }, { true })
+        player.prepare(listOf(sentence(10, "First."), sentence(20, "Second.")), 0)
+        player.play()
+        runCurrent()
+        player.skip(1)
+        runCurrent()
+        assertEquals(listOf(2), played)
+        assertEquals(20, player.state.value.sentenceIndex)
+        player.stop()
+        runCurrent()
+    }
+
+    @Test
+    fun voiceChangeCancelsPlaybackAndRestartsCurrentSentence() = runTest {
+        val voices = mutableListOf<String>()
+        var cancelledAudio = 0
+        val player = DesktopReadAloudPlayer(this, { _, voice -> voices += voice; audio(1) }, {
+            try { awaitCancellation() } finally { cancelledAudio++ }
+        }, { true })
+        player.prepare(listOf(sentence(3, "Hello.")), 0)
+        player.play()
+        runCurrent()
+        player.selectVoice("M1")
+        runCurrent()
+        assertEquals(listOf("F1", "M1"), voices)
+        assertEquals(1, cancelledAudio)
+        assertEquals(3, player.state.value.sentenceIndex)
+        player.stop()
+        runCurrent()
+        assertEquals(2, cancelledAudio)
+    }
+
+    @Test
+    fun lookaheadDoesNotAdvanceHighlightBeforePlayback() = runTest {
+        val synthesized = mutableListOf<String>()
+        val player = DesktopReadAloudPlayer(this, { text, _ -> synthesized += text; audio(1) }, {
+            awaitCancellation()
+        }, { true })
+        player.prepare(listOf(sentence(4, "First."), sentence(8, "Second."), sentence(12, "Third.")), 0)
+        player.play()
+        runCurrent()
+        assertEquals(listOf("First.", "Second."), synthesized)
+        assertEquals(4, player.state.value.sentenceIndex)
+        player.stop()
+        runCurrent()
+    }
+
+    @Test
+    fun fullQueueUsesSpokenTextAndFinishes() = runTest {
+        val synthesized = mutableListOf<String>()
+        val player = DesktopReadAloudPlayer(this, { text, _ -> synthesized += text; audio(1) }, { delay(10) }, { true })
+        player.prepare(listOf(sentence(0, "1").copy(spokenText = "One.", pauseBeforeMillis = 20), sentence(1, "Two.")), 0)
+        player.play()
+        advanceUntilIdle()
+        assertEquals(listOf("One.", "Two."), synthesized)
+        assertFalse(player.state.value.playing)
+        assertFalse(player.state.value.active)
+    }
+
+    @Test
+    fun accentStageCanBeComparedWithPlainEngineAndNeverChangesOriginalText() = runTest {
+        val inputs = mutableListOf<String>()
+        val original = sentence(0, "замок")
+        val player = DesktopReadAloudPlayer(this, { text, _ -> inputs += text; audio(1) }, {}, { true }, { "з+амок" })
+        player.prepare(listOf(original), 0)
+        player.play()
+        advanceUntilIdle()
+        assertEquals(listOf("за\u0301мок"), inputs)
+        assertEquals("замок", original.spokenText)
+        player.selectEngine("supertonic-3")
+        player.play()
+        advanceUntilIdle()
+        assertEquals(listOf("за\u0301мок", "замок"), inputs)
+    }
+
+    @Test
+    fun accentedTextIsRechunkedToFitTtsWithoutDiscardingMarks() = runTest {
+        val inputs = mutableListOf<String>()
+        val marked = "а\u0301 ".repeat(110).trim()
+        val player = DesktopReadAloudPlayer(this, { text, _ -> inputs += text; audio(1) }, {}, { true }, { marked })
+        player.prepare(listOf(sentence(0, "а ".repeat(110))), 0)
+        player.play()
+        advanceUntilIdle()
+        assertTrue(inputs.size > 1)
+        assertTrue(inputs.all { it.length <= 300 })
+        assertEquals(marked.replace(" ", ""), inputs.joinToString("").replace(" ", ""))
+    }
+
+    @Test
+    fun missingModelAndInferenceErrorsRemainVisibleAndRetryable() = runTest {
+        var installed = false
+        var calls = 0
+        val player = DesktopReadAloudPlayer(this, { _, _ -> calls++; error("bad model") }, {}, { installed })
+        player.prepare(listOf(sentence(0, "Hello.")), 0)
+        player.play()
+        assertEquals(0, calls)
+        assertFalse(player.state.value.playing)
+        assertNotNull(player.settingsState.value.message)
+        installed = true
+        player.play()
+        advanceUntilIdle()
+        assertEquals(1, calls)
+        assertTrue(player.state.value.active)
+        assertFalse(player.state.value.buffering)
+        assertNotNull(player.settingsState.value.message)
+    }
+}
diff --git a/composeApp/src/jvmTest/kotlin/net/sergeych/toread/tts/SupertonicTextTest.kt b/composeApp/src/jvmTest/kotlin/net/sergeych/toread/tts/SupertonicTextTest.kt
new file mode 100644
index 0000000..e3ee93a
--- /dev/null
+++ b/composeApp/src/jvmTest/kotlin/net/sergeych/toread/tts/SupertonicTextTest.kt
@@ -0,0 +1,45 @@
+package net.sergeych.toread.tts
+
+import kotlin.test.Test
+import kotlin.test.assertEquals
+import kotlin.test.assertTrue
+
+class SupertonicTextTest {
+    @Test
+    fun longSentencesAndUnbrokenWordsAreBoundedWithoutLosingText() {
+        for (text in listOf("Слово ".repeat(150), "a".repeat(601), "a".repeat(299) + "😀" + "b".repeat(301), "а".repeat(300) + "\u0301" + "а".repeat(50))) {
+            val chunks = SupertonicTts.chunks(text)
+            assertTrue(chunks.all { it.length <= 300 && it.isNotBlank() })
+            assertEquals(text.filterNot(Char::isWhitespace), chunks.joinToString("").filterNot(Char::isWhitespace))
+            assertTrue(chunks.none { Character.isHighSurrogate(it.last()) || Character.isLowSurrogate(it.first()) })
+            assertTrue(chunks.none { it.first() == '\u0301' })
+        }
+    }
+
+    @Test
+    fun languageSelectionHandlesRussianEnglishAndMixedText() {
+        assertEquals("ru", SupertonicTts.languageOf("Ёж шёл домой."))
+        assertEquals("en", SupertonicTts.languageOf("The library."))
+        assertEquals("na", SupertonicTts.languageOf("Он сказал Hello."))
+    }
+
+    @Test
+    fun preprocessingDoesNotInsertPredictedStress() {
+        assertEquals("<ru>На двери висит замок.</ru>", SupertonicTts.prepareText("На двери висит замок", "ru"))
+        assertEquals("<en>Hello!</en>", SupertonicTts.prepareText("Hello !", "en"))
+    }
+
+    @Test
+    fun ruaccentStressBecomesCombiningAcuteOnCyrillicVowels() {
+        assertEquals("за\u0301мок замо\u0301к Ё\u0301лка", SupertonicTts.russianStressToUnicode("з+амок зам+ок +Ёлка"))
+        assertEquals("2+2 +a +б а\u0301", SupertonicTts.russianStressToUnicode("2+2 +a +б а\u0301"))
+        assertEquals("<ru>За\u0301мок.</ru>", SupertonicTts.prepareText("З+амок", "ru"))
+        assertEquals("<ru>За\u0301мок.</ru>", SupertonicTts.prepareText("За\u0301мок", "ru"))
+    }
+
+    @Test
+    fun pcmConversionClipsAndHandlesNonFiniteSamples() {
+        val bytes = SpeechAudio(floatArrayOf(-2f, 2f, Float.NaN), 44100).pcm16()
+        assertEquals(listOf(1, 128, 255, 127, 0, 0), bytes.map { it.toInt() and 255 })
+    }
+}
diff --git a/docs/desktop-tts.md b/docs/desktop-tts.md
new file mode 100644
index 0000000..118bde6
--- /dev/null
+++ b/docs/desktop-tts.md
@@ -0,0 +1,176 @@
+# Desktop CPU speech trial
+
+Desktop Read aloud defaults to **Silero v5_5_ru / Xenia** for Russian when its
+local runtime is installed. Silero runs in a persistent CPU-only Python process
+and handles its own stress and ё restoration. English and mixed-script chunks
+use Supertonic 3 / F1 through ONNX Runtime's CPU provider. Mixed passages retain
+Supertonic's Russian pronunciation limitations. Plain Supertonic is retained as an audition baseline. Initial listening feedback
+found its Russian word stress unacceptable. An optional RUAccent CPU stage now
+adds stress and restores ё before synthesis; its +vowel notation is converted to
+a Cyrillic vowel followed by U+0301. Manual minimal-pair listening showed stress
+changes, but this is not a comprehensive pronunciation-quality evaluation.
+Android retains its existing system TTS implementation.
+
+## Run
+
+From the repository root:
+
+```sh
+python3 scripts/download-tts-model.py
+./gradlew :composeApp:run
+```
+
+The script downloads about 401 MB into the ignored `.local/tts/supertonic-3`
+directory. Downloads are revision-pinned and verified against the committed file
+manifest. It reuses valid files and replaces incomplete downloads only after hash
+verification. No Python packages are needed. The model license is saved alongside
+its weights.
+
+Open a book and choose **Read aloud** from the reader menu. Playback highlights the
+current sentence, follows the reading position, and exposes Stop, previous/next
+sentence, and voice settings. Silero offers Xenia (default), Baya, Aidar, Kseniya
+and Eugene. The Supertonic engines offer F1–F5 and M1–M5. Engine and voice choices
+last for this application session; each launch starts with Xenia if installed.
+
+The model loads lazily on first playback. One chunk is prepared ahead while audio
+plays. Stop, skipping, and voice changes cancel the old playback and lookahead;
+Supertonic inference checks cancellation between model stages/denoising steps;
+cancelling Silero inference terminates its worker, which restarts on demand. Long
+sentences are split at whitespace into at most 300-character chunks. English and
+Russian are selected from their scripts; mixed-script chunks use the model's
+language-agnostic `na` mode. This heuristic is not general language detection.
+
+For a model installed elsewhere, set `YEREED_TTS_MODEL_DIR` (or JVM property
+`yereed.tts.modelDir`) to the directory containing `onnx/` and `voice_styles/`.
+Without an override the app searches parent directories for the workspace model,
+then falls back to `~/.toread/tts/supertonic-3`. Packaged applications can use that
+home-directory location. The application itself never downloads models or sends
+text over the network. The first Gradle build and model setup require downloads.
+
+## Optional RUAccent stage (desktop trial)
+
+```sh
+python3 scripts/setup-ruaccent.py
+./gradlew :composeApp:run
+```
+
+This creates a private Python 3.12 runtime under `.local/ruaccent-runtime` and
+fetches about 726 MB of revision-pinned RUAccent assets into `.local/tts/ruaccent`.
+It does not change system Python. The bootstrap supports Linux x86_64 directly;
+on other systems install `uv` first. The dependency versions and model hashes
+are recorded under `scripts/`. This desktop trial uses RUAccent 1.5.8.3,
+`turbo3.1`, the full dictionary/rule pipeline, and CPU execution with four threads.
+This is a desktop quality experiment, not an Android footprint recommendation.
+
+After setup, restart the reader. **Supertonic 3 + RUAccent** becomes the default
+engine; select **Supertonic 3** in read-aloud settings for a direct comparison.
+English bypasses the accentor. Already marked Russian chunks are preserved.
+The accentor is a persistent local child process communicating over stdin/stdout;
+Hugging Face and Transformers run in offline mode. Book contents are never sent
+to a server. Set `YEREED_RUACCENT_HOME` to override its installation directory.
+Stop/skip kills a pending accentor request when necessary; the next request
+restarts the worker. Displayed book text and stored files are unchanged.
+
+The model recognizes U+0301, and preprocessing preserves it. For example,
+`з+амок` becomes `за́мок`, not a Latin accented-vowel substitution. Predicted
+stress can still be wrong; model-token support does not guarantee compliance.
+
+## Audition / benchmark
+
+```sh
+./gradlew :composeApp:ttsSample
+# Also play the four samples through the desktop audio device:
+./gradlew :composeApp:ttsSample -PttsPlay=true
+# Controlled manual stress pairs, same words/noise seed:
+./gradlew :composeApp:ttsSample -PttsStress=true -PttsPlay=true
+# Original plain Russian passage, then RUAccent + Supertonic:
+./gradlew :composeApp:ttsSample -PttsRuAccent=true -PttsPlay=true
+```
+
+WAV files and their exact unaccented texts go to `composeApp/build/tts-samples/`.
+The sample task uses the same inference and audio code as the reader without
+opening or changing the library. It prints model-load time, synthesis time,
+audio duration, and real-time factor (generation seconds / audio seconds).
+
+Initial measurements on an Intel Core Ultra 9 275HX, four CPU threads, eight
+steps, speed 1.05: 1.075 seconds model load; English/Russian F1/M1 samples generated
+11.6–15.3 seconds of audio in 2.4–3.3 seconds (RTF 0.20–0.24). These are short
+local samples, not phone/battery or long-book benchmarks. Synthesis completion
+and non-silent waveforms do not establish pronunciation quality.
+
+## Dependencies and provenance
+
+- [Supertonic source](https://github.com/supertone-oss-archive/supertonic): the
+  Kotlin inference code adapts the MIT Java example; attribution is in
+  [licenses/supertonic-code-MIT.txt](licenses/supertonic-code-MIT.txt).
+- [Supertonic 3 model](https://huggingface.co/supertone-oss-archive/supertonic-3):
+  revision `aafc6e32416a594460b32413efc49d7fe4ce6d46`, OpenRAIL-M weights.
+  Upstream development/support ended in September 2026.
+- ONNX Runtime 1.23.1, standard CPU artifact; no CUDA/GPU provider.
+- JSON loading uses Jackson; PCM playback uses Java Sound.
+
+Run `./gradlew :composeApp:jvmTest` for text-boundary, PCM conversion, queue,
+lookahead, cancellation, voice-change, and missing-model/error tests, together
+with existing desktop/common tests. Model inference is exercised separately by
+`ttsSample`, so ordinary tests do not download weights or use speakers.
+
+The isolated virtual-display capture did not provide a usable UI image; rendered
+GUI verification remains incomplete. No further screen access is used.
+
+RUAccent measurements on the same host: about 5.9 seconds cold load, 30–43 ms
+for the two warm short passages, and about 2.1 GiB resident memory for the full
+accentor worker alone. The combined passage took 2.98 seconds of TTS synthesis
+for 16.42 seconds of audio, excluding accentor startup. In that passage RUAccent
+incorrectly chose `за́мок` for the door lock; on a shorter explicit door-lock
+sentence it chose `замо́к`. These are limited observations, not an accuracy score.
+
+## Silero Russian audition
+
+User listening feedback rejected Supertonic's Russian quality. A separate CPU
+audition uses official Silero `v5_5_ru`, with its own stress, ё, stress-homograph
+and ё-homograph processing enabled. It does not use RUAccent. The same model is
+now the preferred desktop Russian engine. The Russian model is CC BY-NC-SA 4.0, unlike Silero's MIT CIS
+base models; see [upstream](https://github.com/snakers4/silero-models).
+
+Setup from the repository root (requires `uv`; our existing local copy is
+`.local/tools/uv`):
+
+```sh
+export UV_CACHE_DIR="$PWD/.local/uv-cache"
+export UV_PYTHON_INSTALL_DIR="$PWD/.local/python"
+export UV_SYSTEM_CERTS=true
+.local/tools/uv venv --python 3.12.14 .local/silero-runtime
+.local/tools/uv pip sync --python .local/silero-runtime/bin/python scripts/silero-requirements.txt --index-url https://download.pytorch.org/whl/cpu
+mkdir -p .local/tts/silero-v5_5-ru
+curl -fL https://models.silero.ai/models/tts/ru/v5_5_ru.pt -o .local/tts/silero-v5_5-ru/model.pt.part
+mv .local/tts/silero-v5_5-ru/model.pt.part .local/tts/silero-v5_5-ru/model.pt
+.local/silero-runtime/bin/python scripts/silero-sample.py
+python3 - <<'PY'
+from pathlib import Path
+import json
+root = Path.cwd()
+(root / '.local/tts/silero-v5_5-ru/runtime.json').write_text(json.dumps({
+    'python': str(root / '.local/silero-runtime/bin/python')
+}) + '\n')
+PY
+# Restart the reader after setup, or check its actual Kotlin/Python bridge:
+./gradlew :composeApp:ttsSample -PttsSilero=true
+```
+
+Outputs: `composeApp/build/tts-samples/silero-*.wav`, matching input `.txt`
+files, and `silero-benchmark.json` with timings, settings and the downloaded
+model's SHA-256. The script uses CPU-only PyTorch, four threads and 48 kHz PCM.
+Default voices are Xenia, Baya and Aidar; `--voices xenia` limits the audition.
+Each reads the earlier library passage and a separate unaccented ё/context
+challenge. Pronunciation quality must be judged by listening.
+
+The desktop integration discovers `silero-v5_5-ru` beside the Supertonic model
+directory; `YEREED_SILERO_HOME` can override it. `runtime.json` points to the
+isolated Python executable. The worker script is bundled in the application;
+model loading and inference are local. Both models are needed for bilingual use.
+
+Initial audition on the same desktop: model load 0.38 seconds; six samples
+generated 13.2–15.5 seconds of audio in 0.36–0.82 seconds (RTF 0.027–0.059,
+including the first synthesis). All outputs were finite and non-silent.
+Downloaded model SHA-256:
+`50081637b602126ee06cb3bc8a744d25651d2da149ee8864b9a379bfdd934437`.
diff --git a/docs/licenses/supertonic-code-MIT.txt b/docs/licenses/supertonic-code-MIT.txt
new file mode 100644
index 0000000..943171b
--- /dev/null
+++ b/docs/licenses/supertonic-code-MIT.txt
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Supertone Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
\ No newline at end of file
diff --git a/gradle/libs.versions.toml b/gradle/libs.versions.toml
index 472124e..87ffcca 100644
--- a/gradle/libs.versions.toml
+++ b/gradle/libs.versions.toml
@@ -19,8 +19,12 @@ ktor = "3.4.3"
 logback = "1.5.32"
 material3 = "1.11.0-alpha07"
 h2 = "2.4.240"
+onnxruntime = "1.23.1"
+jackson = "2.18.6"
 
 [libraries]
+onnxruntime = { module = "com.microsoft.onnxruntime:onnxruntime", version.ref = "onnxruntime" }
+jackson-databind = { module = "com.fasterxml.jackson.core:jackson-databind", version.ref = "jackson" }
 kotlin-test = { module = "org.jetbrains.kotlin:kotlin-test", version.ref = "kotlin" }
 kotlin-testJunit = { module = "org.jetbrains.kotlin:kotlin-test-junit", version.ref = "kotlin" }
 junit = { module = "junit:junit", version.ref = "junit" }
@@ -41,6 +45,7 @@ compose-components-resources = { module = "org.jetbrains.compose.components:comp
 compose-uiToolingPreview = { module = "org.jetbrains.compose.ui:ui-tooling-preview", version.ref = "composeMultiplatform" }
 kotlinx-coroutinesSwing = { module = "org.jetbrains.kotlinx:kotlinx-coroutines-swing", version.ref = "kotlinx-coroutines" }
 kotlinx-coroutinesCore = { module = "org.jetbrains.kotlinx:kotlinx-coroutines-core", version.ref = "kotlinx-coroutines" }
+kotlinx-coroutinesTest = { module = "org.jetbrains.kotlinx:kotlinx-coroutines-test", version.ref = "kotlinx-coroutines" }
 h2 = { module = "com.h2database:h2", version.ref = "h2" }
 logback = { module = "ch.qos.logback:logback-classic", version.ref = "logback" }
 ktor-serverCore = { module = "io.ktor:ktor-server-core-jvm", version.ref = "ktor" }
diff --git a/scripts/download-tts-model.py b/scripts/download-tts-model.py
new file mode 100644
index 0000000..309f766
--- /dev/null
+++ b/scripts/download-tts-model.py
@@ -0,0 +1,56 @@
+#!/usr/bin/env python3
+"""Download the pinned Supertonic 3 CPU model. No packages or credentials required."""
+import hashlib
+import json
+from pathlib import Path
+import sys
+import urllib.request
+
+ROOT = Path(__file__).resolve().parent.parent
+MANIFEST = json.loads((ROOT / "scripts/supertonic-3-model.json").read_text())
+
+
+def valid(path, entry):
+    if not path.is_file() or path.stat().st_size != entry["size"]:
+        return False
+    digest = hashlib.sha256() if "sha256" in entry else hashlib.sha1()
+    if "git_blob" in entry:
+        digest.update(f"blob {entry['size']}\0".encode())
+    with path.open("rb") as source:
+        for chunk in iter(lambda: source.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest() == entry.get("sha256", entry.get("git_blob"))
+
+
+def download(manifest, destination):
+    total = sum(entry["size"] for entry in manifest["files"])
+    print(f"{manifest['repository']}: {total / 1e6:.1f} MB; destination: {destination}", flush=True)
+    base = f"https://huggingface.co/{manifest['repository']}/resolve/{manifest['revision']}"
+    for entry in manifest["files"]:
+        target = destination / entry["path"]
+        if valid(target, entry):
+            print(f"Verified {entry['path']}", flush=True)
+            continue
+        target.parent.mkdir(parents=True, exist_ok=True)
+        temporary = target.with_suffix(target.suffix + ".part")
+        print(f"Downloading {entry['path']} ({entry['size'] / 1e6:.1f} MB)", flush=True)
+        try:
+            with urllib.request.urlopen(f"{base}/{entry['path']}", timeout=60) as source, temporary.open("wb") as output:
+                while chunk := source.read(1024 * 1024):
+                    output.write(chunk)
+            if not valid(temporary, entry):
+                raise RuntimeError(f"Checksum mismatch: {entry['path']}")
+            temporary.replace(target)
+        finally:
+            temporary.unlink(missing_ok=True)
+
+
+def main():
+    destination = Path(sys.argv[1]) if len(sys.argv) > 1 else ROOT / ".local/tts/supertonic-3"
+    print("Supertonic model license: OpenRAIL-M (saved as LICENSE alongside the model).", flush=True)
+    download(MANIFEST, destination)
+    print("Ready. Run ./gradlew :composeApp:run and use Read aloud in a book.", flush=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/ruaccent-model.json b/scripts/ruaccent-model.json
new file mode 100644
index 0000000..87eaae7
--- /dev/null
+++ b/scripts/ruaccent-model.json
@@ -0,0 +1,231 @@
+{
+  "repository": "ruaccent/accentuator",
+  "revision": "b78ae5ea1e62beaf138bed1865cd8c3b0b5ca855",
+  "files": [
+    {
+      "path": "README.md",
+      "size": 2660,
+      "git_blob": "7f12645cf20ed21742fd63a3b799efd8c7007734"
+    },
+    {
+      "path": "dictionary/accents.json.gz",
+      "size": 20954156,
+      "sha256": "aa460ebba90de00fbbf3d41d121961f605b98667e45efb7920f127473b15515e"
+    },
+    {
+      "path": "dictionary/accents_nn.json.gz",
+      "size": 845996,
+      "sha256": "8395664000b80c1afe09bfea3650945b0933482b8e3dee5bb9d429eb18c44935"
+    },
+    {
+      "path": "dictionary/omographs.json.gz",
+      "size": 219047,
+      "sha256": "04a9e81c68d65f65ba493fe0110f99e79087548c2beeec3032e2b66e28706f36"
+    },
+    {
+      "path": "dictionary/rule_engine/accents.json",
+      "size": 694782,
+      "git_blob": "91690536e14e133fe4017941a466fde663b96011"
+    },
+    {
+      "path": "dictionary/rule_engine/forms.json",
+      "size": 2669,
+      "git_blob": "6686cd615fb55c09976550aeba3ad6687ecc32dd"
+    },
+    {
+      "path": "dictionary/yo_homographs.json.gz",
+      "size": 5747,
+      "sha256": "c4ee777bbbab87f9eac838f370ad92974e079d02b21903e480c54b5f0c8c60d1"
+    },
+    {
+      "path": "dictionary/yo_words.json.gz",
+      "size": 548914,
+      "sha256": "a19fa89a964a0691d9fe4ee384783e3934904891843d8f59a1c480d67947a82a"
+    },
+    {
+      "path": "koziev/rulemma/rulemma.dat",
+      "size": 16703198,
+      "sha256": "bf2b3ef3ff7a0aa6e4250aa4e9c8ed568e25f825deebdb12dee1b46b785ba9fc"
+    },
+    {
+      "path": "koziev/rulemma/rulemma.py",
+      "size": 11166,
+      "git_blob": "a182b7d1694b13e8ac6807e2701ad17bf4064bbe"
+    },
+    {
+      "path": "koziev/rupostagger/__init__.py",
+      "size": 111,
+      "git_blob": "01bd02ccfc68c851542390a6d0a684481f897bb9"
+    },
+    {
+      "path": "koziev/rupostagger/database/ruword2tags.db",
+      "size": 168816640,
+      "sha256": "a06848e656bef642aafb4440c03554fa78f2f32dde92ea66f3f86ce9977b167e"
+    },
+    {
+      "path": "koziev/rupostagger/rupostagger.config",
+      "size": 276,
+      "git_blob": "4794511fc1b5ef4bc649fab5eafdf6b775206b72"
+    },
+    {
+      "path": "koziev/rupostagger/rupostagger.model",
+      "size": 2417464,
+      "sha256": "21b7b0bfd7427b5fdc1604052176db8aa3b139b3ce03be440cfce48536f8e5ef"
+    },
+    {
+      "path": "koziev/rupostagger/rupostagger.py",
+      "size": 7118,
+      "git_blob": "198a8db45678a9578dfad5bbf01d3efd8e95a7c8"
+    },
+    {
+      "path": "koziev/rupostagger/rusyllab.py",
+      "size": 18975,
+      "git_blob": "7f606a48ff2a5de6b6388a45caf1d41bd6089611"
+    },
+    {
+      "path": "koziev/rupostagger/ruword2tags.dat",
+      "size": 9683765,
+      "sha256": "dde47b5f1d48ff899887ac07812dcabd2966e48e84646f3065bfd06627c2af58"
+    },
+    {
+      "path": "koziev/rupostagger/ruword2tags.py",
+      "size": 18086,
+      "git_blob": "71098637747f19e44788506775e4cd1e6c5e7183"
+    },
+    {
+      "path": "nn/nn_accent/big.onnx",
+      "size": 2285217,
+      "sha256": "47e69d9ae19f2a82e21b1c70f6a4bbfb1abc5759e98b2e67d009c5e9d7af18c9"
+    },
+    {
+      "path": "nn/nn_accent/config.json",
+      "size": 841,
+      "git_blob": "0d1656a1fb93fcea726186c8bfad977c6e27c4a5"
+    },
+    {
+      "path": "nn/nn_accent/model.onnx",
+      "size": 803402,
+      "sha256": "4e393144e45626f6f1062a0784ef06f921b97321a8e7b87ac2a09a892286500a"
+    },
+    {
+      "path": "nn/nn_accent/ort_config.json",
+      "size": 727,
+      "git_blob": "5a8fc726323dda820ae4e794e4d0d514de07eee2"
+    },
+    {
+      "path": "nn/nn_accent/special_tokens_map.json",
+      "size": 99,
+      "git_blob": "1543e951b7c6f9c246a5e441d7162b2c65ab48a5"
+    },
+    {
+      "path": "nn/nn_accent/tokenizer_config.json",
+      "size": 257,
+      "git_blob": "5ea10565e634adcd37089652fbf6bb2cf05c1530"
+    },
+    {
+      "path": "nn/nn_accent/vocab.txt",
+      "size": 140,
+      "git_blob": "1e5a3d765a6f5a686ed3c6294efb85e9e00a431a"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/added_tokens.json",
+      "size": 279418,
+      "git_blob": "23620bf92f2ad9e1243202cff127617fe9415c9a"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/config.json",
+      "size": 723,
+      "git_blob": "2ff7dbeb01fffcb05f65ad430b08205fe6668693"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/merges.txt",
+      "size": 1213606,
+      "git_blob": "2d4658a914a3760fbf1a3a9c72b42eefcf8d8ded"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/model.onnx",
+      "size": 359306923,
+      "sha256": "2cb6a174c4cdb45bd3132b4f7c8a3779fc4b6869863180ed7d0e421bcd453dbd"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/special_tokens_map.json",
+      "size": 280,
+      "git_blob": "d5698132694f4f1bcff08fa7d937b1701812598e"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/tokenizer.json",
+      "size": 5532343,
+      "git_blob": "31ac709e3c4625c6bf80b7c97066a9cde4d9ba89"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/tokenizer_config.json",
+      "size": 492,
+      "git_blob": "15a19568f93d1b292c7ff6d4ef0fdac13bd65b53"
+    },
+    {
+      "path": "nn/nn_omograph/turbo3.1/vocab.json",
+      "size": 1555282,
+      "git_blob": "ede65202ec9ef483ff166bb31f76615f30048db7"
+    },
+    {
+      "path": "nn/nn_stress_usage_predictor/config.json",
+      "size": 822,
+      "git_blob": "df189246d87953bead249c2ee82e3985a96be8f8"
+    },
+    {
+      "path": "nn/nn_stress_usage_predictor/model.onnx",
+      "size": 116473561,
+      "sha256": "3d547500637b4ddfec8880ed6d1405fd50ee9d3f0131ef8a2a69dcf961dbefeb"
+    },
+    {
+      "path": "nn/nn_stress_usage_predictor/special_tokens_map.json",
+      "size": 125,
+      "git_blob": "a8b3208c2884c4efb86e49300fdd3dc877220cdf"
+    },
+    {
+      "path": "nn/nn_stress_usage_predictor/tokenizer.json",
+      "size": 2413536,
+      "git_blob": "b4d5ccba4a193e0330537378bb0bba47026bd364"
+    },
+    {
+      "path": "nn/nn_stress_usage_predictor/tokenizer_config.json",
+      "size": 368,
+      "git_blob": "d7d40c5e58b2c6ba58974571f617abe23221eb6c"
+    },
+    {
+      "path": "nn/nn_stress_usage_predictor/vocab.txt",
+      "size": 1080667,
+      "git_blob": "211012a925fc51581997b6d3ef047bcda9a842f4"
+    },
+    {
+      "path": "nn/nn_yo_homograph_resolver/config.json",
+      "size": 625,
+      "git_blob": "6615961d500dfe1fbbf302d1bbd4832b7e27f398"
+    },
+    {
+      "path": "nn/nn_yo_homograph_resolver/model.onnx",
+      "size": 14332169,
+      "sha256": "42cc85bf0c4b319dfe3d89fa17b162a92fd5e1c651a657cb3d5f44978d4e70ac"
+    },
+    {
+      "path": "nn/nn_yo_homograph_resolver/special_tokens_map.json",
+      "size": 125,
+      "git_blob": "a8b3208c2884c4efb86e49300fdd3dc877220cdf"
+    },
+    {
+      "path": "nn/nn_yo_homograph_resolver/tokenizer.json",
+      "size": 126952,
+      "git_blob": "bbab1a4bc4dfa68697338718621489e77d03e70a"
+    },
+    {
+      "path": "nn/nn_yo_homograph_resolver/tokenizer_config.json",
+      "size": 401,
+      "git_blob": "7d8a1ab4371909b9dcaabfcf93b9bbd56045605a"
+    },
+    {
+      "path": "nn/nn_yo_homograph_resolver/vocab.txt",
+      "size": 48576,
+      "git_blob": "32bc54d43a46178e70dcefb85626f52e9ddbc98b"
+    }
+  ]
+}
diff --git a/scripts/ruaccent-requirements.txt b/scripts/ruaccent-requirements.txt
new file mode 100644
index 0000000..c0065b3
--- /dev/null
+++ b/scripts/ruaccent-requirements.txt
@@ -0,0 +1,25 @@
+certifi==2026.7.22
+charset-normalizer==3.5.1
+filelock==4.0.4
+flatbuffers==25.12.19
+fsspec==2026.9.0
+hf-xet==1.6.0
+huggingface-hub==0.36.2
+idna==3.20
+numpy==2.5.3
+onnxruntime==1.30.0
+packaging==26.3
+protobuf==7.36.2
+python-crfsuite==0.9.12
+pyyaml==6.0.3
+razdel==0.5.0
+regex==2026.9.10
+requests==2.34.2
+ruaccent==1.5.8.3
+safetensors==0.8.0
+sentencepiece==0.2.2
+tokenizers==0.22.2
+tqdm==4.70.1
+transformers==4.57.6
+typing-extensions==4.16.0
+urllib3==2.8.0
diff --git a/scripts/setup-ruaccent.py b/scripts/setup-ruaccent.py
new file mode 100644
index 0000000..711c997
--- /dev/null
+++ b/scripts/setup-ruaccent.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+"""Workspace-local RuAccent CPU setup. No sudo or changes to system Python."""
+import hashlib
+import io
+import json
+import os
+from pathlib import Path
+import platform
+import runpy
+import shutil
+import subprocess
+import urllib.request
+import zipfile
+
+ROOT = Path(__file__).resolve().parent.parent
+
+
+def main():
+    uv = shutil.which("uv") or str(ROOT / ".local/tools/uv")
+    if not Path(uv).is_file():
+        if platform.system() != "Linux" or platform.machine() != "x86_64":
+            raise SystemExit("Install uv first on this platform, then rerun this setup script.")
+        with urllib.request.urlopen("https://pypi.org/pypi/uv/0.12.19/json", timeout=60) as source:
+            metadata = json.load(source)
+        wheel = next(x for x in metadata["urls"] if x["filename"].endswith("manylinux_2_17_x86_64.manylinux2014_x86_64.whl"))
+        with urllib.request.urlopen(wheel["url"], timeout=60) as source:
+            content = source.read()
+        expected = "a63d18a0aa38ee9f21a5406afbbaeb41303bcd954be9d6b7c1b95ac275e53958"
+        if hashlib.sha256(content).hexdigest() != expected:
+            raise RuntimeError("uv checksum mismatch")
+        with zipfile.ZipFile(io.BytesIO(content)) as archive:
+            name = next(x for x in archive.namelist() if x.endswith("/uv"))
+            Path(uv).parent.mkdir(parents=True, exist_ok=True)
+            Path(uv).write_bytes(archive.read(name))
+        Path(uv).chmod(0o755)
+    environment = dict(os.environ, UV_CACHE_DIR=str(ROOT / ".local/uv-cache"),
+                       UV_PYTHON_INSTALL_DIR=str(ROOT / ".local/python"), UV_SYSTEM_CERTS="true")
+    venv = ROOT / ".local/ruaccent-runtime"
+    python = venv / ("Scripts/python.exe" if os.name == "nt" else "bin/python")
+    if not python.is_file():
+        subprocess.run([uv, "venv", "--python", "3.12.14", str(venv)], env=environment, check=True)
+    subprocess.run([uv, "pip", "sync", "--python", str(python), str(ROOT / "scripts/ruaccent-requirements.txt")], env=environment, check=True)
+    model = ROOT / ".local/tts/ruaccent"
+    manifest = json.loads((ROOT / "scripts/ruaccent-model.json").read_text())
+    runpy.run_path(str(ROOT / "scripts/download-tts-model.py"))["download"](manifest, model)
+    # RUAccent 1.5.8.3 expects its rule-engine files beside the installed package.
+    module_path = subprocess.check_output([str(python), "-c", "import importlib.util; print(importlib.util.find_spec('ruaccent').submodule_search_locations[0])"], text=True).strip()
+    shutil.copytree(model / "koziev", Path(module_path) / "koziev", dirs_exist_ok=True)
+    worker = ROOT / "composeApp/src/jvmMain/resources/tts/ruaccent-worker.py"
+    shutil.copyfile(worker, model / "worker.py")
+    # Verify a wholly offline initialization before declaring the installation ready.
+    subprocess.run([str(python), str(model / "worker.py"), str(model), "--check"], check=True, env=environment)
+    (model / "runtime.json").write_text(json.dumps({"python": str(python), "model": "turbo3.1", "revision": manifest["revision"]}, indent=2) + "\n")
+    print("RuAccent ready: CPU, turbo3.1, full dictionary; no network during inference.", flush=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/silero-requirements.txt b/scripts/silero-requirements.txt
new file mode 100644
index 0000000..988591e
--- /dev/null
+++ b/scripts/silero-requirements.txt
@@ -0,0 +1,10 @@
+filelock==3.32.3
+fsspec==2026.7.0
+jinja2==3.1.6
+markupsafe==3.0.3
+mpmath==1.3.0
+networkx==3.6.1
+setuptools==78.1.0
+sympy==1.14.0
+torch==2.8.0+cpu
+typing-extensions==4.16.0
diff --git a/scripts/silero-sample.py b/scripts/silero-sample.py
new file mode 100644
index 0000000..0558aaa
--- /dev/null
+++ b/scripts/silero-sample.py
@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+"""Audition official Silero v5_5_ru locally on CPU, without changing the reader."""
+import argparse
+import array
+import hashlib
+import json
+import os
+from pathlib import Path
+import sys
+import time
+import wave
+
+ROOT = Path(__file__).resolve().parent.parent
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+os.environ.setdefault("TORCH_HOME", str(ROOT / ".local/torch"))
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--voices", nargs="+", default=["xenia", "baya", "aidar"])
+    args = parser.parse_args()
+    import torch
+
+    torch.set_num_threads(4)
+    torch.set_num_interop_threads(1)
+    torch.manual_seed(42)
+    model_path = ROOT / ".local/tts/silero-v5_5-ru/model.pt"
+    output = ROOT / "composeApp/build/tts-samples"
+    output.mkdir(parents=True, exist_ok=True)
+    start = time.perf_counter()
+    model = torch.package.PackageImporter(str(model_path)).load_pickle("tts_models", "model")
+    model.to(torch.device("cpu"))
+    loaded = time.perf_counter() - start
+    print(f"Silero v5_5_ru, CPU, 4 threads; load {loaded:.3f}s; voices {model.speakers}", flush=True)
+    samples = {
+        "ru": "В библиотеке было тихо. Она открыла книгу и начала читать. За окном шёл осенний дождь. На двери висел старый замок, а вдали виднелся замок на холме. Завтра, подумала она, мы наконец доберёмся до моря.",
+        "ru-yo": "Елка росла у дома. Ежик шел по лесу. Ребенок пьет теплое молоко. Все уже пришли. Он прочитал все. Она все еще ждет. Он узнает об этом завтра. Он узнает ее по голосу.",
+    }
+    report = {"model": "v5_5_ru", "sha256": hashlib.sha256(model_path.read_bytes()).hexdigest(),
+              "torch": torch.__version__, "device": "cpu", "threads": 4, "load_seconds": loaded,
+              "put_accent": True, "put_yo": True, "put_stress_homo": True, "put_yo_homo": True, "samples": []}
+    for voice in args.voices:
+        for label, text in samples.items():
+            start = time.perf_counter()
+            with torch.inference_mode():
+                audio = model.apply_tts(text=text, speaker=voice, sample_rate=48000,
+                                        put_accent=True, put_yo=True,
+                                        put_stress_homo=True, put_yo_homo=True).detach().cpu().flatten()
+            elapsed = time.perf_counter() - start
+            assert audio.numel() and torch.isfinite(audio).all() and audio.abs().max() > 0.001
+            duration = audio.numel() / 48000
+            pcm = array.array("h", (audio.clamp(-1, 1) * 32767).round().to(torch.int16).tolist())
+            if sys.byteorder != "little":
+                pcm.byteswap()
+            path = output / f"silero-{label}-{voice}.wav"
+            with wave.open(str(path), "wb") as wav:
+                wav.setparams((1, 2, 48000, 0, "NONE", "not compressed"))
+                wav.writeframes(pcm.tobytes())
+            path.with_suffix(".txt").write_text(text + "\n")
+            report["samples"].append({"file": path.name, "synthesis_seconds": elapsed,
+                                      "audio_seconds": duration, "rtf": elapsed / duration})
+            print(f"{path.name}: {elapsed:.3f}s synthesis, {duration:.3f}s audio, RTF {elapsed / duration:.3f}", flush=True)
+    (output / "silero-benchmark.json").write_text(json.dumps(report, indent=2) + "\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/supertonic-3-model.json b/scripts/supertonic-3-model.json
new file mode 100644
index 0000000..c2e8584
--- /dev/null
+++ b/scripts/supertonic-3-model.json
@@ -0,0 +1,91 @@
+{
+  "repository": "supertone-oss-archive/supertonic-3",
+  "revision": "aafc6e32416a594460b32413efc49d7fe4ce6d46",
+  "files": [
+    {
+      "path": "LICENSE",
+      "size": 15007,
+      "git_blob": "1e2cbe6cef94c8cdf3ed1fcebc0f5317ca7ad5a1"
+    },
+    {
+      "path": "onnx/duration_predictor.onnx",
+      "size": 3700147,
+      "sha256": "c3eb91414d5ff8a7a239b7fe9e34e7e2bf8a8140d8375ffb14718b1c639325db"
+    },
+    {
+      "path": "onnx/text_encoder.onnx",
+      "size": 36416150,
+      "sha256": "c7befd5ea8c3119769e8a6c1486c4edc6a3bc8365c67621c881bbb774b9902ff"
+    },
+    {
+      "path": "onnx/tts.json",
+      "size": 8253,
+      "git_blob": "28575b0793c92607a0f0a292550df9bd17709a79"
+    },
+    {
+      "path": "onnx/unicode_indexer.json",
+      "size": 277676,
+      "git_blob": "3aaf4d33024aff0be455633c24b3636b3a810150"
+    },
+    {
+      "path": "onnx/vector_estimator.onnx",
+      "size": 256534781,
+      "sha256": "883ac868ea0275ef0e991524dc64f16b3c0376efd7c320af6b53f5b780d7c61c"
+    },
+    {
+      "path": "onnx/vocoder.onnx",
+      "size": 101424195,
+      "sha256": "085de76dd8e8d5836d6ca66826601f615939218f90e519f70ee8a36ed2a4c4ba"
+    },
+    {
+      "path": "voice_styles/F1.json",
+      "size": 292046,
+      "git_blob": "421365f307ed1535d8da16845031d0e9fa3e60c5"
+    },
+    {
+      "path": "voice_styles/F2.json",
+      "size": 292423,
+      "git_blob": "b09a3a433df8a4856d1d7d64736a56ddf37a4ea3"
+    },
+    {
+      "path": "voice_styles/F3.json",
+      "size": 290794,
+      "git_blob": "a366f94c77c4b440404071f1a52bed3ccf83a5ff"
+    },
+    {
+      "path": "voice_styles/F4.json",
+      "size": 291808,
+      "git_blob": "39c78a52795320c6101b7eaf984c0e2352fbdc18"
+    },
+    {
+      "path": "voice_styles/F5.json",
+      "size": 291479,
+      "git_blob": "06983f8816ad00a8eeb7efd37a33d457587e16f8"
+    },
+    {
+      "path": "voice_styles/M1.json",
+      "size": 291748,
+      "git_blob": "bddfff85a7c4c140c11620bd7005d86681dccdac"
+    },
+    {
+      "path": "voice_styles/M2.json",
+      "size": 292055,
+      "git_blob": "602ff6c4d6d48cdb71c37461f7c175d71fe8b34b"
+    },
+    {
+      "path": "voice_styles/M3.json",
+      "size": 290198,
+      "git_blob": "000b60436b77619b933fdc33628ad737595e6794"
+    },
+    {
+      "path": "voice_styles/M4.json",
+      "size": 291522,
+      "git_blob": "aae253ae81185fa06a6469c3f1cdff04ed121ddf"
+    },
+    {
+      "path": "voice_styles/M5.json",
+      "size": 291469,
+      "git_blob": "0536b252698830f8536c077edcb94e1ce4768ab6"
+    }
+  ]
+}