@goodandready/dsh-tts 0.3.22 → 0.3.23

package/lib/client.js

@@ -154,24 +154,47 @@ window.__ModuleLoader__.load({
       })
     }
 
+    function prepareAudioItem(item) {
+      if (!item || !item.audioBase64 || item.prepared) return item
+      try {
+        const bin = atob(item.audioBase64)
+        const bytes = new Uint8Array(bin.length)
+        for (let i = 0; i < bin.length; i++) bytes[i] = bin.charCodeAt(i)
+        const blob = new Blob([bytes], { type: item.mime || 'audio/mpeg' })
+        const url = URL.createObjectURL(blob)
+        const audio = new Audio(url)
+        audio.preload = 'auto'
+        item.prepared = { audio, url }
+      } catch (err) { /* decode error */ }
+      return item
+    }
+
+    function prebufferNext() {
+      for (const next of player.queue) {
+        if (next && next.audioBase64 && !next.prepared) {
+          prepareAudioItem(next)
+          break
+        }
+      }
+    }
+
     function playAudio(item) {
       return new Promise((resolve) => {
         try {
-          const bin = atob(item.audioBase64)
-          const bytes = new Uint8Array(bin.length)
-          for (let i = 0; i < bin.length; i++) bytes[i] = bin.charCodeAt(i)
-          const blob = new Blob([bytes], { type: item.mime || 'audio/mpeg' })
-          const url = URL.createObjectURL(blob)
-          const audio = new Audio(url)
+          prepareAudioItem(item)
+          if (!item.prepared) return resolve()
+          const { audio, url } = item.prepared
           audio.playbackRate = Math.max(0.5, Math.min(2, player.rate || 1))
           player.audio = audio
           playerChanged()
+          prebufferNext()
           let doneCalled = false
           const done = () => {
             if (doneCalled) return
             doneCalled = true
             currentAudioResolve = null
-            URL.revokeObjectURL(url)
+            try { URL.revokeObjectURL(url) } catch (_) {}
+            item.prepared = null
             player.audio = null
             playerChanged()
             resolve()
@@ -211,6 +234,12 @@ window.__ModuleLoader__.load({
     }
 
     function stopPlayback() {
+      for (const item of player.queue) {
+        if (item && item.prepared && item.prepared.url) {
+          try { URL.revokeObjectURL(item.prepared.url) } catch (_) {}
+          item.prepared = null
+        }
+      }
       player.queue.length = 0
       player.paused = false
       if (currentAudioResolve) {

package/lib/index.js

@@ -309,6 +309,9 @@ export function apply(ctx, config) {
   // Счётчики синтеза: в памяти, обнуляются рестартом и DELETE /stats.
   const stats = { total: 0, cacheHits: 0, errors: 0, providers: {} }
 
+  // Дедупликация параллельных запросов на синтез (защита от Cache Stampede)
+  const inFlightSyntheses = new Map()
+
   // Кэш синтеза живёт рядом с прочими данными харнесса.
   const speechCache = createSpeechCache({
     root: process.env.DSH_HOME || path.join(os.homedir(), '.dsh'),
@@ -403,20 +406,34 @@ export function apply(ctx, config) {
         }
       }
     }
-    const providers = makeProviders(
-      { resolveKey, fetchImpl: fetch, cfg, modelManager },
-      { text, lang: cfg.autoDetect ? detectLang(text) : cfg.language, signal, models, voices, role },
-    )
-    const out = await runChain(order, providers)
-    stats.total++
-    const ps = stats.providers[out.provider] || (stats.providers[out.provider] = { n: 0, e: 0, ms: 0 })
-    ps.n++
-    ps.ms += out.tookMs || 0
-    if (cacheAllowed) {
-      const p = out.provider
-      speechCache.put(cacheKey([text, p, models[p], voices[p], role, ov.ssmlStyle]), out.mime, out.audio).catch(() => {})
+    const flightKey = cacheKey([text, order.join(','), JSON.stringify(models), JSON.stringify(voices), role, ov.ssmlStyle])
+    if (inFlightSyntheses.has(flightKey)) {
+      return inFlightSyntheses.get(flightKey)
     }
-    return out
+
+    const synthPromise = (async () => {
+      try {
+        const providers = makeProviders(
+          { resolveKey, fetchImpl: fetch, cfg, modelManager },
+          { text, lang: cfg.autoDetect ? detectLang(text) : cfg.language, signal, models, voices, role },
+        )
+        const out = await runChain(order, providers)
+        stats.total++
+        const ps = stats.providers[out.provider] || (stats.providers[out.provider] = { n: 0, e: 0, ms: 0 })
+        ps.n++
+        ps.ms += out.tookMs || 0
+        if (cacheAllowed) {
+          const p = out.provider
+          speechCache.put(cacheKey([text, p, models[p], voices[p], role, ov.ssmlStyle]), out.mime, out.audio).catch(() => {})
+        }
+        return out
+      } finally {
+        inFlightSyntheses.delete(flightKey)
+      }
+    })()
+
+    inFlightSyntheses.set(flightKey, synthPromise)
+    return synthPromise
   }
 
   // Синтез одной фразы с постановкой в очередь. Порядок сохраняется: номер

package/lib/providers.js

@@ -75,12 +75,32 @@ function asBuffer(body) {
 
 export function makeProviders(deps, req) {
   const { resolveKey, fetchImpl, cfg } = deps
-  const { text, lang, signal, models, voices } = req
+  const { text, lang, signal: parentSignal, models, voices } = req
+  const cloudTimeoutMs = Number(cfg.cloudTimeoutMs) > 0 ? Number(cfg.cloudTimeoutMs) : 10000
+
+  function withCloudTimeout() {
+    const controller = new AbortController()
+    let timedOut = false
+    const timer = setTimeout(() => {
+      timedOut = true
+      controller.abort(new Error(`Cloud TTS timeout after ${cloudTimeoutMs}ms`))
+    }, cloudTimeoutMs)
+    if (parentSignal) {
+      if (parentSignal.aborted) controller.abort(parentSignal.reason)
+      else parentSignal.addEventListener('abort', () => controller.abort(parentSignal.reason), { once: true })
+    }
+    return {
+      signal: controller.signal,
+      cleanup: () => clearTimeout(timer),
+    }
+  }
 
   async function openai() {
     const key = await resolveKey(cfg.openaiKeyEnv)
     if (!key) return { ok: false, provider: 'openai', reason: `no ${cfg.openaiKeyEnv}` }
-    const res = await fetchImpl((cfg.openaiBaseUrl || 'https://api.openai.com/v1') + '/audio/speech', {
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl((cfg.openaiBaseUrl || 'https://api.openai.com/v1') + '/audio/speech', {
       method: 'POST',
       headers: { authorization: `Bearer ${key}`, 'content-type': 'application/json' },
       body: JSON.stringify({
@@ -94,6 +114,9 @@ export function makeProviders(deps, req) {
     if (!res.ok) throw new Error(`OpenAI HTTP ${res.status}`)
     const audio = asBuffer(Buffer.from(await res.arrayBuffer()))
     return { ok: audio.length > 0, provider: 'openai', audio, mime: 'audio/mpeg', reason: audio.length ? '' : 'empty audio' }
+    } finally {
+      cleanup()
+    }
   }
 
   // Одна реализация на всех, кто говорит на языке OpenAI: адрес и имя ключа
@@ -106,7 +129,9 @@ export function makeProviders(deps, req) {
       if (!token) return { ok: false, provider: key, reason: `no ${keyEnv}` }
       const base = cfg[key + 'BaseUrl'] || spec.baseUrl
       if (!base) return { ok: false, provider: key, reason: 'no base url configured' }
-      const res = await fetchImpl(base + '/audio/speech', {
+      const { signal, cleanup } = withCloudTimeout()
+      try {
+        const res = await fetchImpl(base + '/audio/speech', {
         method: 'POST',
         headers: { authorization: `Bearer ${token}`, 'content-type': 'application/json' },
         body: JSON.stringify({
@@ -120,6 +145,9 @@ export function makeProviders(deps, req) {
       if (!res.ok) throw new Error(`${key} HTTP ${res.status}`)
       const audio = asBuffer(Buffer.from(await res.arrayBuffer()))
       return { ok: audio.length > 0, provider: key, audio, mime: 'audio/mpeg', reason: audio.length ? '' : 'empty audio' }
+      } finally {
+        cleanup()
+      }
     }
   }
 
@@ -130,7 +158,9 @@ export function makeProviders(deps, req) {
     if (!key) return { ok: false, provider: 'mimo', reason: `no ${cfg.mimoKeyEnv}` }
     const base = (cfg.mimoBaseUrl || 'https://api.xiaomimimo.com/v1').replace(/\/+$/, '')
     const format = cfg.mimoFormat === 'wav' ? 'wav' : 'mp3'
-    const res = await fetchImpl(base + '/chat/completions', {
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl(base + '/chat/completions', {
       method: 'POST',
       headers: { authorization: `Bearer ${key}`, 'content-type': 'application/json', accept: 'application/json' },
       body: JSON.stringify({
@@ -160,6 +190,9 @@ export function makeProviders(deps, req) {
       mime: format === 'wav' ? 'audio/wav' : 'audio/mpeg',
       reason: audio.length ? '' : 'empty audio',
     }
+    } finally {
+      cleanup()
+    }
   }
 
   // MiniMax — через их утилиту: своего протокола здесь нет намеренно.
@@ -181,7 +214,9 @@ export function makeProviders(deps, req) {
     const key = await resolveKey(cfg.elevenlabsKeyEnv)
     if (!key) return { ok: false, provider: 'elevenlabs', reason: `no ${cfg.elevenlabsKeyEnv}` }
     const voice = pick(DEFAULT_VOICES, voices, 'elevenlabs')
-    const res = await fetchImpl(`https://api.elevenlabs.io/v1/text-to-speech/${encodeURIComponent(voice)}`, {
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl(`https://api.elevenlabs.io/v1/text-to-speech/${encodeURIComponent(voice)}`, {
       method: 'POST',
       headers: { 'xi-api-key': key, 'content-type': 'application/json', accept: 'audio/mpeg' },
       body: JSON.stringify({
@@ -193,6 +228,9 @@ export function makeProviders(deps, req) {
     if (!res.ok) throw new Error(`ElevenLabs HTTP ${res.status}`)
     const audio = asBuffer(Buffer.from(await res.arrayBuffer()))
     return { ok: audio.length > 0, provider: 'elevenlabs', audio, mime: 'audio/mpeg', reason: audio.length ? '' : 'empty audio' }
+    } finally {
+      cleanup()
+    }
   }
 
   async function google() {
@@ -200,7 +238,9 @@ export function makeProviders(deps, req) {
     if (!key) return { ok: false, provider: 'google', reason: `no ${cfg.googleKeyEnv}` }
     const model = pick(DEFAULT_MODELS, models, 'google')
     const voiceName = pick(DEFAULT_VOICES, voices, 'google')
-    const res = await fetchImpl(
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl(
       `https://generativelanguage.googleapis.com/v1beta/models/${encodeURIComponent(model)}:generateContent?key=${encodeURIComponent(key)}`,
       {
         method: 'POST',
@@ -221,6 +261,9 @@ export function makeProviders(deps, req) {
     if (!b64) return { ok: false, provider: 'google', reason: 'no inline audio' }
     const audio = Buffer.from(b64, 'base64')
     return { ok: audio.length > 0, provider: 'google', audio, mime: 'audio/wav', reason: audio.length ? '' : 'empty audio' }
+    } finally {
+      cleanup()
+    }
   }
 
   async function azure() {
@@ -238,7 +281,9 @@ export function makeProviders(deps, req) {
     const ssml = `<speak version="1.0" xml:lang="${lang || 'en-US'}">`
       + `<voice xml:lang="${lang || 'en-US'}" name="${voice}">`
       + `${inner}</voice></speak>`
-    const res = await fetchImpl(`https://${region}.tts.speech.microsoft.com/cognitiveservices/v1`, {
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl(`https://${region}.tts.speech.microsoft.com/cognitiveservices/v1`, {
       method: 'POST',
       headers: {
         'Ocp-Apim-Subscription-Key': key,
@@ -251,12 +296,17 @@ export function makeProviders(deps, req) {
     if (!res.ok) throw new Error(`Azure HTTP ${res.status}`)
     const audio = asBuffer(Buffer.from(await res.arrayBuffer()))
     return { ok: audio.length > 0, provider: 'azure', audio, mime: 'audio/mpeg', reason: audio.length ? '' : 'empty audio' }
+    } finally {
+      cleanup()
+    }
   }
 
   async function groq() {
     const key = await resolveKey(cfg.groqKeyEnv)
     if (!key) return { ok: false, provider: 'groq', reason: `no ${cfg.groqKeyEnv}` }
-    const res = await fetchImpl('https://api.groq.com/openai/v1/audio/speech', {
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl('https://api.groq.com/openai/v1/audio/speech', {
       method: 'POST',
       headers: { authorization: `Bearer ${key}`, 'content-type': 'application/json' },
       body: JSON.stringify({
@@ -270,13 +320,18 @@ export function makeProviders(deps, req) {
     if (!res.ok) throw new Error(`Groq HTTP ${res.status}`)
     const audio = asBuffer(Buffer.from(await res.arrayBuffer()))
     return { ok: audio.length > 0, provider: 'groq', audio, mime: 'audio/mpeg', reason: audio.length ? '' : 'empty audio' }
+    } finally {
+      cleanup()
+    }
   }
 
   async function deepgram() {
     const key = await resolveKey(cfg.deepgramKeyEnv)
     if (!key) return { ok: false, provider: 'deepgram', reason: `no ${cfg.deepgramKeyEnv}` }
     const model = pick(DEFAULT_MODELS, models, 'deepgram')
-    const res = await fetchImpl(`https://api.deepgram.com/v1/speak?model=${encodeURIComponent(model)}`, {
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl(`https://api.deepgram.com/v1/speak?model=${encodeURIComponent(model)}`, {
       method: 'POST',
       headers: { authorization: `Token ${key}`, 'content-type': 'application/json' },
       body: JSON.stringify({ text }),
@@ -285,12 +340,17 @@ export function makeProviders(deps, req) {
     if (!res.ok) throw new Error(`Deepgram HTTP ${res.status}`)
     const audio = asBuffer(Buffer.from(await res.arrayBuffer()))
     return { ok: audio.length > 0, provider: 'deepgram', audio, mime: 'audio/mpeg', reason: audio.length ? '' : 'empty audio' }
+    } finally {
+      cleanup()
+    }
   }
 
   async function openrouter() {
     const key = await resolveKey(cfg.openrouterKeyEnv)
     if (!key) return { ok: false, provider: 'openrouter', reason: `no ${cfg.openrouterKeyEnv}` }
-    const res = await fetchImpl('https://openrouter.ai/api/v1/audio/speech', {
+    const { signal, cleanup } = withCloudTimeout()
+    try {
+      const res = await fetchImpl('https://openrouter.ai/api/v1/audio/speech', {
       method: 'POST',
       headers: { authorization: `Bearer ${key}`, 'content-type': 'application/json' },
       body: JSON.stringify({
@@ -304,6 +364,9 @@ export function makeProviders(deps, req) {
     if (!res.ok) throw new Error(`OpenRouter HTTP ${res.status}`)
     const audio = asBuffer(Buffer.from(await res.arrayBuffer()))
     return { ok: audio.length > 0, provider: 'openrouter', audio, mime: 'audio/mpeg', reason: audio.length ? '' : 'empty audio' }
+    } finally {
+      cleanup()
+    }
   }
 
   async function edge() {
@@ -340,6 +403,9 @@ export function makeProviders(deps, req) {
     try {
       const modelPath = deps.modelManager ? deps.modelManager.getModelPath('kokoro') : (cfg.kokoroModelPath || '')
       const engine = new KokoroEngine({ modelPath })
+      if (!engine.isInstalled()) {
+        return { ok: false, provider: 'kokoro', reason: 'Kokoro ONNX model not installed. Use Piper or eSpeak for offline TTS.' }
+      }
       const voice = pick(DEFAULT_VOICES, voices, 'kokoro') || 'af_bella'
       const audio = await engine.synthesizeWav(text, voice)
       return { ok: audio && audio.length > 0, provider: 'kokoro', audio, mime: 'audio/wav' }
@@ -353,8 +419,11 @@ export function makeProviders(deps, req) {
       const modelPath = deps.modelManager ? deps.modelManager.getModelPath('f5') : (cfg.f5ModelPath || '')
       const engine = new F5Engine({ modelPath })
       const ok = await engine.ping()
-      if (!ok) return { ok: false, provider: 'f5', reason: 'F5 daemon unavailable' }
+      if (!ok) return { ok: false, provider: 'f5', reason: 'F5 daemon unavailable. Use Piper or eSpeak for offline TTS.' }
       const kokoroFallback = new KokoroEngine({ modelPath })
+      if (!kokoroFallback.isInstalled()) {
+        return { ok: false, provider: 'f5', reason: 'F5 model weights not installed. Use Piper or eSpeak for offline TTS.' }
+      }
       const audio = await kokoroFallback.synthesizeWav(text, 'af_bella')
       return { ok: true, provider: 'f5', audio, mime: 'audio/wav' }
     } catch (e) {

package/lib/text.js

@@ -147,6 +147,32 @@ export function assistantText(message) {
   return blocksToText(message.content)
 }
 
+// Защита распространенных сокращений от ложного разбиения на предложения
+const PROTECTED_ABBREVIATIONS = [
+  // Русские сокращения
+  'т.е.', 'т.д.', 'т.п.', 'т.к.', 'руб.', 'коп.', 'ул.', 'пер.', 'пр.', 'просп.',
+  'д.', 'корп.', 'стр.', 'кв.', 'г.', 'гг.', 'в.', 'вв.', 'см.', 'ср.', 'напр.',
+  'рис.', 'табл.', 'им.', 'акад.', 'проф.', 'доц.', 'млн.', 'млрд.', 'тыс.',
+  // Английские сокращения
+  'e.g.', 'i.e.', 'vs.', 'etc.', 'dr.', 'mr.', 'mrs.', 'ms.', 'prof.',
+  'approx.', 'dept.', 'est.', 'min.', 'max.', 'no.', 'vol.', 'p.m.', 'a.m.'
+]
+
+export function protectAbbreviations(text) {
+  let res = String(text || '')
+  for (const abbr of PROTECTED_ABBREVIATIONS) {
+    const escaped = abbr.replace(/\./g, '\\.')
+    const rx = new RegExp('(?<=^|[\\s"«(])' + escaped + '(?=[\\s"»)\\]]|$)', 'gi')
+    res = res.replace(rx, (m) => m.replace(/\./g, '\u2024'))
+  }
+  res = res.replace(/(?<=[vV]?\d+)\.(?=\d+)/g, '\u2024')
+  return res
+}
+
+export function restoreAbbreviations(text) {
+  return String(text || '').replace(/\u2024/g, '.')
+}
+
 /**
  * Режет текст на куски по границам предложений.
  *
@@ -161,9 +187,10 @@ export function assistantText(message) {
 export function splitSentences(text, maxChars) {
   const limit = Number(maxChars) > 0 ? Number(maxChars) : 320
   const min = Math.min(60, Math.floor(limit / 3))
-  const source = String(text || '').trim()
-  if (!source) return []
+  const raw = String(text || '').trim()
+  if (!raw) return []
 
+  const source = protectAbbreviations(raw)
   const pieces = []
   let current = ''
   // Границей считаем точку, вопрос, восклицание, многоточие и перевод строки.
@@ -181,5 +208,5 @@ export function splitSentences(text, maxChars) {
     }
   }
   if (current) pieces.push(current)
-  return pieces.filter(Boolean)
+  return pieces.filter(Boolean).map(restoreAbbreviations)
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-tts",
-  "version": "0.3.22",
+  "version": "0.3.23",
   "description": "Text-to-speech for DeepSeek Harness: speak agent replies in the Web UI with a provider fallback chain (OpenAI, ElevenLabs, Google, Azure, Groq, Deepgram, OpenRouter, Edge, Piper, eSpeak).",
   "license": "MIT",
   "type": "module",
@@ -63,4 +63,4 @@
   "publishConfig": {
     "access": "public"
   }
-}
+}

package/lib/client/worklet.js

@@ -1,71 +0,0 @@
-class TTSWorklet extends AudioWorkletProcessor {
-  constructor() {
-    super()
-    this.bufferQueue = []
-    this.currentChunk = null
-    this.chunkOffset = 0
-    this.playing = true
-
-    this.port.onmessage = (event) => {
-      const data = event.data
-      if (!data) return
-      if (data.type === 'chunk' && data.pcm instanceof Float32Array) {
-        this.bufferQueue.push(data.pcm)
-      } else if (data.type === 'clear' || data.type === 'stop') {
-        this.bufferQueue = []
-        this.currentChunk = null
-        this.chunkOffset = 0
-      } else if (data.type === 'pause') {
-        this.playing = false
-      } else if (data.type === 'resume') {
-        this.playing = true
-      }
-    }
-  }
-
-  process(inputs, outputs, parameters) {
-    const output = outputs[0]
-    if (!output || !output[0]) return true
-    const channel = output[0]
-    const bufferLength = channel.length
-
-    if (!this.playing) {
-      channel.fill(0)
-      return true
-    }
-
-    let written = 0
-    while (written < bufferLength) {
-      if (!this.currentChunk || this.chunkOffset >= this.currentChunk.length) {
-        if (this.bufferQueue.length === 0) {
-          this.currentChunk = null
-          break
-        }
-        this.currentChunk = this.bufferQueue.shift()
-        this.chunkOffset = 0
-      }
-
-      const available = this.currentChunk.length - this.chunkOffset
-      const needed = bufferLength - written
-      const toCopy = Math.min(available, needed)
-
-      for (let i = 0; i < toCopy; i++) {
-        channel[written + i] = this.currentChunk[this.chunkOffset + i]
-      }
-
-      this.chunkOffset += toCopy
-      written += toCopy
-    }
-
-    // Fill remaining output buffer with silence if queue is drained
-    if (written < bufferLength) {
-      for (let i = written; i < bufferLength; i++) {
-        channel[i] = 0
-      }
-    }
-
-    return true
-  }
-}
-
-registerProcessor('tts-worklet', TTSWorklet)

package/lib/router.js

@@ -1,4 +0,0 @@
-export class AdaptiveRouter {
-  constructor({registry, role, lang}){ this.registry=registry; this.role=role; this.lang=lang }
-  pick(){ const list=this.registry.list(); return list[0]?.id || null }
-}