@goodandready/dsh-voice 0.8.5 → 0.8.6

package/lib/client.js

@@ -96,6 +96,11 @@ window.__ModuleLoader__.load({
       'vocabulary': 'Custom vocabulary (one word per line)',
       'polish': 'Polish transcript with model',
       'polishHint': 'Fix punctuation and fillers via the harness model before inserting',
+      'stream': 'Continuous dictation',
+      'streamHint': 'Cut phrases by a timer while you speak instead of waiting for a long pause',
+      'streamChunkMs': 'Stream chunk (ms)',
+      'vadAdapt': 'Adaptive silence',
+      'vadAdaptHint': 'Auto-tune the silence threshold to the pace of your speech. 0 = fixed',
       'voiceCommandsLabel': 'Voice edit commands ("new line", "paragraph")',
       'normalizeTranscriptHint': 'transcribe_audio: spoken numbers to digits, tidy punctuation',
       'messageTitle': 'Voice message',
@@ -189,6 +194,11 @@ window.__ModuleLoader__.load({
       'vocabulary': 'Свой словарь (одно слово в строке)',
       'polish': 'Полировка текста моделью',
       'polishHint': 'Пунктуация и слова-паразиты через модель харнесса перед вставкой',
+      'stream': 'Непрерывная диктовка',
+      'streamHint': 'Резать фразы по таймеру во время речи, а не по длинной паузе',
+      'streamChunkMs': 'Кусок потока (мс)',
+      'vadAdapt': 'Адаптивная тишина',
+      'vadAdaptHint': 'Авто-подстройка порога под темп речи. 0 = фикс. поведение',
       'voiceCommandsLabel': 'Голосовые команды («с новой строки», «абзац»)',
       'normalizeTranscriptHint': 'transcribe_audio: числа словами — в цифры, аккуратная пунктуация',
       'messageTitle': 'Голосовое сообщение',
@@ -247,7 +257,7 @@ window.__ModuleLoader__.load({
       pending: null,      // {text, leftMs} — окно отмены режима message
       inputActions: null,
       input: null,
-      settings: { vadSilenceMs: 700, autoSendMs: 4000 },
+      settings: { vadSilenceMs: 700, autoSendMs: 4000, stream: false, streamChunkMs: 1200, vadAdapt: 0 },
       listeners: new Set(),
       notify() { this.listeners.forEach((l) => l()) },
       set(patch) { Object.assign(this, patch); this.notify() },
@@ -535,6 +545,7 @@ window.__ModuleLoader__.load({
         audioCtx: null, analyser: null,
         cutting: false, closing: false,
         silenceMs: 0, hadSpeech: false,
+        streamMs: 0,
       }
       recorder.ondataavailable = (e) => { if (e.data && e.data.size > 0) rec.chunks.push(e.data) }
       const AC = typeof AudioContext !== 'undefined' ? AudioContext
@@ -573,6 +584,7 @@ window.__ModuleLoader__.load({
         rec.chunks = []
         rec.silenceMs = 0
         rec.hadSpeech = false
+        rec.streamMs = 0
         if (!rec.closing) {
           try { rec.recorder.start() } catch (e) { /* поток закрылся */ }
         }
@@ -833,7 +845,29 @@ window.__ModuleLoader__.load({
           else if (rec.hadSpeech) rec.silenceMs += tick
           const speaking = level > 0.06
           if (voice.speaking !== speaking) { voice.speaking = speaking; voice.notify() }
-          if (rec.mode === 'dictation' && rec.hadSpeech && rec.silenceMs >= voice.settings.vadSilenceMs) {
+          const adapt = Number(voice.settings.vadAdapt) || 0
+          let effectiveVad = Number(voice.settings.vadSilenceMs) || 700
+          if (adapt > 0 && rec.hadSpeech) {
+            // Адаптивный порог (#41): считаем плотность речи за последние
+            // ~1s (20 сэмплов). Плотная речь -> порог ниже (точнее режем),
+            // паузная -> порог растёт к базе (не режем на вдохе).
+            const win = voice.levels.slice(-20)
+            const density = win.length ? win.filter((v) => v > 0.06).length / win.length : 0
+            const k = adapt * (density - 0.5) * 2
+            effectiveVad = Math.max(150, Math.round(Number(voice.settings.vadSilenceMs) * (1 - k)))
+          }
+          // Непрерывный режим (#40): режем по таймеру непрерывной речи,
+          // не дожидаясь длинной паузы.
+          const stream = !!voice.settings.stream && rec.mode === 'dictation'
+          if (stream && rec.hadSpeech && !rec.cutting) {
+            rec.streamMs += tick
+            const chunk = Number(voice.settings.streamChunkMs) || 1200
+            if (rec.streamMs >= chunk) { cutPhrase(); return }
+          } else {
+            rec.streamMs = 0
+          }
+          if (rec.mode === 'dictation' && rec.hadSpeech && rec.silenceMs >= effectiveVad) {
+            rec.streamMs = 0
             cutPhrase()
           }
         }, tick)
@@ -973,6 +1007,9 @@ window.__ModuleLoader__.load({
                 historyLimit: Number(data && data.historyLimit),
                 voiceCommands: !!(data && data.voiceCommands),
                 sendDelayMs: Number(data && data.modes && data.modes.dictation && data.modes.dictation.sendDelayMs) || 0,
+                stream: !!(data && data.modes && data.modes.dictation && data.modes.dictation.stream),
+                streamChunkMs: Number(data && data.modes && data.modes.dictation && data.modes.dictation.streamChunkMs) || 1200,
+                vadAdapt: Number(data && data.modes && data.modes.dictation && data.modes.dictation.vadAdapt) || 0,
               })
             })
             .catch(() => { /* без подсказки хоста клавиши просто не будет */ })
@@ -1227,6 +1264,9 @@ window.__ModuleLoader__.load({
           historyLimit: Number(snap && snap.historyLimit),
           voiceCommands: !!(snap && snap.voiceCommands),
           sendDelayMs: Number(value && value.dictation && value.dictation.sendDelayMs) || 0,
+          stream: !!(value && value.dictation && value.dictation.stream),
+          streamChunkMs: Number(value && value.dictation && value.dictation.streamChunkMs) || 1200,
+          vadAdapt: Number(value && value.dictation && value.dictation.vadAdapt) || 0,
         })
       }, [ready, value, snap])
 
@@ -1274,6 +1314,9 @@ window.__ModuleLoader__.load({
           historyLimit: Number(draft.historyLimit),
           voiceCommands: !!draft.voiceCommands,
           sendDelayMs: Number(draft.dictation && draft.dictation.sendDelayMs) || 0,
+          stream: !!(draft.dictation && draft.dictation.stream),
+          streamChunkMs: Number(draft.dictation && draft.dictation.streamChunkMs) || 1200,
+          vadAdapt: Number(draft.dictation && draft.dictation.vadAdapt) || 0,
         })
         // Композер держит обработчик клавиши: пусть перечитает настройку,
         // иначе новая клавиша заработает только после перезагрузки страницы.
@@ -1370,6 +1413,18 @@ window.__ModuleLoader__.load({
               onChange: (e) => setIn('dictation', 'polish', e.target.checked),
             })),
           React.createElement('div', { className: 'dvs-sub' }, t('polishHint')),
+          React.createElement('label', { className: 'dvs-field', title: t('streamHint') }, t('stream'),
+            React.createElement('input', {
+              type: 'checkbox', checked: !!(draft && draft.dictation && draft.dictation.stream), disabled: !writable,
+              onChange: (e) => setIn('dictation', 'stream', e.target.checked),
+            })),
+          numField('dictation', 'streamChunkMs', t('streamChunkMs'), t('streamHint')),
+          React.createElement('label', { className: 'dvs-field', title: t('vadAdaptHint') }, t('vadAdapt'),
+            React.createElement('input', {
+              type: 'range', min: 0, max: 1, step: 0.1,
+              value: Number(draft && draft.dictation && draft.dictation.vadAdapt) || 0, disabled: !writable,
+              onChange: (e) => setIn('dictation', 'vadAdapt', Number(e.target.value)),
+            })),
         ),
         React.createElement('div', { className: 'dvs-block' },
           React.createElement('div', { className: 'dvs-h' }, t('messageTitle')),

package/lib/index.js

@@ -69,6 +69,12 @@ export const Config = z.object({
       .description('Dictation: wait this many ms after a phrase before appending it, with a cancel window. 0 disables the delay.'),
     polish: z.boolean().default(false)
       .description('Polish the transcript through the harness model before inserting: punctuation, paragraphs, filler-word removal.'),
+    stream: z.boolean().default(false)
+      .description('Continuous dictation: cut phrases by a timer instead of waiting for a long silence, so text flows while you speak.'),
+    streamChunkMs: z.number().default(1200)
+      .description('Continuous dictation: phrase length in ms of uninterrupted speech before the chunk is sent.'),
+    vadAdapt: z.number().min(0).max(1).default(0)
+      .description('Adaptive silence threshold: 0 = fixed (current behaviour); >0 shrinks the threshold during dense speech and grows it during pauses.'),
   }).default({}),
   message: z.object({
     chain: z.array(ChainEntry)
@@ -271,6 +277,8 @@ export function apply(ctx, baseConfig) {
             chain: cfg.dictation.chain, language: cfg.dictation.language,
             vadSilenceMs: cfg.dictation.vadSilenceMs,
             sendDelayMs: cfg.dictation.sendDelayMs, polish: cfg.dictation.polish,
+            stream: cfg.dictation.stream, streamChunkMs: cfg.dictation.streamChunkMs,
+            vadAdapt: cfg.dictation.vadAdapt,
           },
           message: {
             chain: cfg.message.chain, language: cfg.message.language,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-voice",
-  "version": "0.8.5",
+  "version": "0.8.6",
   "description": "Voice input for DeepSeek Harness: dictation chunked by pauses and voice messages, each with its own provider fallback chain (Deepgram, Groq, HuggingFace, local whisper.cpp, plus any OpenAI-compatible API of your own).",
   "license": "MIT",
   "type": "module",