@goodandready/dsh-tts 0.2.0 → 0.3.0

package/lib/client.js

@@ -66,6 +66,7 @@ window.__ModuleLoader__.load({
       rate: 1,
       chime: 'ding',
       queue: [],
+      bargeIn: true,
       busy: false,
       paused: false,
       listeners: new Set(),
@@ -190,6 +191,24 @@ window.__ModuleLoader__.load({
       playerChanged()
     }
 
+    // Человек заговорил — замолкаем.
+    //
+    // Событие приходит из плагина голосового ввода, но никакой связи с ним
+    // нет: если его не установили, событие просто никогда не придёт. Слушаем
+    // окно, а не чужую службу, ровно поэтому.
+    function listenForVoice() {
+      if (typeof window === 'undefined') return () => {}
+      const onSpeaking = (event) => {
+        const phase = event && event.detail && event.detail.phase
+        if (phase !== 'start') return
+        if (!player.bargeIn) return
+        if (!player.audio && !player.queue.length && !player.busy) return
+        stopPlayback()
+      }
+      window.addEventListener('dsh-voice:speaking', onSpeaking)
+      return () => window.removeEventListener('dsh-voice:speaking', onSpeaking)
+    }
+
     async function pollPending() {
       try {
         const st = await fetch('/dsh-tts/status', { cache: 'no-store' })
@@ -198,6 +217,7 @@ window.__ModuleLoader__.load({
         player.enabled = !!(meta && meta.speakReplies)
         if (meta && typeof meta.rate === 'number') player.rate = meta.rate
         if (meta && meta.chime) player.chime = meta.chime
+        if (meta && typeof meta.bargeIn === 'boolean') player.bargeIn = meta.bargeIn
         if (!player.enabled) return
         const res = await fetch('/dsh-tts/pending?after=' + encodeURIComponent(player.after || ''), { cache: 'no-store' })
         if (!res.ok) return
@@ -420,6 +440,8 @@ window.__ModuleLoader__.load({
           boolField('speakAsItGoes', 'Speak as it goes',
             'Read each reply the moment it lands, sentence by sentence, instead of waiting for the whole turn.'),
           numberField('rate', 'Playback speed', '0.5 to 2. Synthesis is untouched; the browser plays faster or slower.', 0.1),
+          boolField('bargeIn', 'Fall silent when you speak',
+            'Stop reading the moment the microphone opens.'),
           boolField('announceApproval', 'Announce approvals',
             'Say it out loud and play a chime when the agent stops for an approval or asks a question.'),
           textField('approvalText', 'Approval wording', 'Said when an approval is asked for; the tool name is appended.'),
@@ -479,6 +501,7 @@ window.__ModuleLoader__.load({
     exports.apply = function apply(ctx) {
       registerSettings(ctx)
       registerSpeaker(ctx)
+      ctx.effect(() => listenForVoice(), 'dsh-tts: замолкать, когда человек заговорил')
       setInterval(pollPending, 1000)
     }
     return module.exports

package/lib/index.js

@@ -45,6 +45,11 @@ export const Config = z.object({
     .number()
     .description('Playback speed, 0.5 to 2. Synthesis is untouched; the browser plays faster or slower.')
     .default(1),
+  bargeIn: z
+    .boolean()
+    .description('Fall silent the moment the microphone opens. Listening to a reply and talking over it '
+      + 'at the same time does not work, and the reply would be recorded along with the voice.')
+    .default(true),
   announceApproval: z
     .boolean()
     .description('Say out loud when the agent stops and waits for an approval, and play a short chime. '
@@ -78,6 +83,13 @@ export const Config = z.object({
   azureKeyEnv: z.string().default('AZURE_SPEECH_KEY'),
   azureRegion: z.string().default('')
     .description('Azure Speech region, e.g. eastus. Required for the azure provider.'),
+  mimoKeyEnv: z.string().description('Credential holding the Xiaomi MiMo key.').default('MIMO_API_KEY'),
+  mimoBaseUrl: z.string().description('MiMo API root. Synthesis goes through its chat endpoint.').default('https://api.xiaomimimo.com/v1'),
+  mimoFormat: z.string().description('MiMo output format: mp3 or wav.').default('mp3'),
+  minimaxBin: z.string()
+    .description('MiniMax CLI, looked up in PATH unless absolute. Install and log in separately; '
+      + 'without it the provider declines and the chain moves on.')
+    .default('mmx'),
   siliconflowKeyEnv: z.string().description('Credential holding the SiliconFlow key.').default('SILICONFLOW_API_KEY'),
   deepinfraKeyEnv: z.string().description('Credential holding the DeepInfra key.').default('DEEPINFRA_API_KEY'),
   fireworksKeyEnv: z.string().description('Credential holding the Fireworks key.').default('FIREWORKS_API_KEY'),
@@ -360,6 +372,7 @@ export function apply(ctx, config) {
         speakReplies: cfg.speakReplies,
         rate: cfg.rate,
         chime: cfg.chime,
+        bargeIn: cfg.bargeIn,
         language: cfg.language,
         chain: cfg.chain,
         providers: PROVIDER_KEYS,

package/lib/local.js

@@ -81,3 +81,30 @@ export async function edgeSpeak(text, { bin, voice, timeoutMs }) {
   }
 }
 
+
+/**
+ * MiniMax через их консольную утилиту.
+ *
+ * Прямого HTTP тут намеренно нет: у MiniMax свой протокол с отдельным
+ * идентификатором группы и своей формой ответа, а официальная утилита уже
+ * умеет и авторизацию, и выбор голоса. Она же хранит вход — плагину не нужно
+ * знать ни ключа, ни устройства их API.
+ *
+ * Утилита ставится отдельно и один раз логинится; без неё провайдер честно
+ * отказывает, и цепочка идёт дальше.
+ */
+export async function minimaxSpeak(text, { bin, voice, model, timeoutMs }) {
+  const dir = await mkdtemp(path.join(os.tmpdir(), 'dsh-tts-minimax-'))
+  const outFile = path.join(dir, 'out.mp3')
+  try {
+    const args = ['speech', 'synthesize', '--text', text, '--out', outFile, '--quiet', '--non-interactive']
+    if (voice) args.push('--voice', voice)
+    if (model) args.push('--model', model)
+    await spawnCollect(bin || 'mmx', args, { timeoutMs })
+    const bytes = await readFile(outFile)
+    if (!bytes || bytes.length < 64) throw new Error('minimax produced empty audio')
+    return { audio: bytes, mime: 'audio/mpeg' }
+  } finally {
+    await rm(dir, { recursive: true, force: true })
+  }
+}

package/lib/providers.js

@@ -1,9 +1,9 @@
-import { edgeSpeak, espeakSpeak, piperSpeak } from './local.js'
+import { edgeSpeak, espeakSpeak, minimaxSpeak, piperSpeak } from './local.js'
 
 export const PROVIDER_KEYS = [
   'openai', 'elevenlabs', 'google', 'azure', 'groq', 'deepgram', 'openrouter',
-  'siliconflow', 'deepinfra', 'fireworks',
-  'edge', 'piper', 'espeak',
+  'siliconflow', 'deepinfra', 'fireworks', 'mimo',
+  'edge', 'piper', 'espeak', 'minimax',
 ]
 
 // Провайдеры, говорящие на языке OpenAI: тот же путь /audio/speech, отличаются
@@ -16,6 +16,8 @@ export const OPENAI_COMPATIBLE = {
 }
 
 export const DEFAULT_MODELS = {
+  mimo: 'mimo-v2.5-tts',
+  minimax: '',
   siliconflow: 'FunAudioLLM/CosyVoice2-0.5B',
   deepinfra: 'hexgrad/Kokoro-82M',
   fireworks: 'kokoro',
@@ -32,6 +34,9 @@ export const DEFAULT_MODELS = {
 }
 
 export const DEFAULT_VOICES = {
+  // Голоса у MiMo названы по-китайски; «冰糖» — их же значение по умолчанию.
+  mimo: '冰糖',
+  minimax: '',
   siliconflow: 'FunAudioLLM/CosyVoice2-0.5B:alex',
   deepinfra: 'af_bella',
   fireworks: 'af_bella',
@@ -106,6 +111,60 @@ export function makeProviders(deps, req) {
     }
   }
 
+  // Xiaomi MiMo. Синтез идёт через чат: текст кладётся сообщением, а голос и
+  // формат — отдельным полем; готовый звук приезжает в ответе base64.
+  async function mimo() {
+    const key = await resolveKey(cfg.mimoKeyEnv)
+    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', {
+      method: 'POST',
+      headers: { authorization: `Bearer ${key}`, 'content-type': 'application/json', accept: 'application/json' },
+      body: JSON.stringify({
+        model: pick(DEFAULT_MODELS, models, 'mimo'),
+        messages: [{ role: 'assistant', content: text }],
+        audio: { format, voice: pick(DEFAULT_VOICES, voices, 'mimo') },
+        stream: false,
+      }),
+      signal,
+    })
+    const data = await res.json().catch(() => ({}))
+    if (!res.ok) {
+      const detail = data && data.error && data.error.message
+      throw new Error(`MiMo HTTP ${res.status}${detail ? ': ' + detail : ''}`)
+    }
+    const encoded = data && data.choices && data.choices[0]
+      && data.choices[0].message && data.choices[0].message.audio
+      && data.choices[0].message.audio.data
+    if (typeof encoded !== 'string' || !encoded) {
+      return { ok: false, provider: 'mimo', reason: 'в ответе нет choices[0].message.audio.data' }
+    }
+    const audio = Buffer.from(encoded, 'base64')
+    return {
+      ok: audio.length > 0,
+      provider: 'mimo',
+      audio,
+      mime: format === 'wav' ? 'audio/wav' : 'audio/mpeg',
+      reason: audio.length ? '' : 'empty audio',
+    }
+  }
+
+  // MiniMax — через их утилиту: своего протокола здесь нет намеренно.
+  async function minimax() {
+    try {
+      const out = await minimaxSpeak(text, {
+        bin: cfg.minimaxBin,
+        voice: pick(DEFAULT_VOICES, voices, 'minimax'),
+        model: pick(DEFAULT_MODELS, models, 'minimax'),
+        timeoutMs: cfg.timeoutMs,
+      })
+      return { ok: true, provider: 'minimax', audio: out.audio, mime: out.mime }
+    } catch (e) {
+      return { ok: false, provider: 'minimax', reason: String(e && e.message || e) }
+    }
+  }
+
   async function elevenlabs() {
     const key = await resolveKey(cfg.elevenlabsKeyEnv)
     if (!key) return { ok: false, provider: 'elevenlabs', reason: `no ${cfg.elevenlabsKeyEnv}` }
@@ -262,7 +321,7 @@ export function makeProviders(deps, req) {
   const compatible = {}
   for (const key of Object.keys(OPENAI_COMPATIBLE)) compatible[key] = openaiCompatible(key)
 
-  return { openai, elevenlabs, google, azure, groq, deepgram, openrouter, ...compatible, edge, piper, espeak }
+  return { openai, elevenlabs, google, azure, groq, deepgram, openrouter, mimo, ...compatible, edge, piper, espeak, minimax }
 }
 
 function escapeXml(s) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-tts",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "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",