@goodandready/dsh-voice 0.9.43 → 0.9.44

package/lib/client.js

@@ -852,8 +852,7 @@ window.__ModuleLoader__.load({
       const text = chunks.join(' ')
       if (!text || text.length < 3) return []
       const matches = text.match(/[A-Za-zА-Яа-яЁё_][A-Za-zА-Яа-яЁё0-9_]{2,29}/g) || []
-      // Linguistic filter: functional stop-word dictionary for vocabulary keyword extraction.
-      // Language processing data for context analysis, distinct from user-facing UI text.
+      // Stop-word dictionary for vocabulary keyword extraction
       const stop = new Set([
         'the', 'and', 'for', 'are', 'but', 'not', 'you', 'all', 'any', 'can', 'her', 'was',
         'one', 'our', 'out', 'day', 'get', 'has', 'him', 'his', 'how', 'man', 'new', 'now',
@@ -983,9 +982,7 @@ window.__ModuleLoader__.load({
       return lead + s + trail
     }
 
-    // Functional speech edit commands (#37): spoken line breaks and punctuation -> symbols.
-    // Linguistic data for audio transcript post-processing; not user-facing UI text.
-    // Russian phrases match RU STT output; English covers EN dictation.
+    // Speech edit commands (#37): spoken line breaks and punctuation -> symbols
     const VOICE_COMMANDS = [
       [/(^|[\s,.!?])с новой строки([\s,.!?]|$)/gi, '$1\n$2'],
       [/(^|[\s,.!?])новая строка([\s,.!?]|$)/gi, '$1\n$2'],
@@ -1033,22 +1030,26 @@ window.__ModuleLoader__.load({
     function insertDraftText(text, target) {
       if (voice.disposed) return false
       const span = target && target.span
+      const sIn = (target && target.input) || voice.input
+      const sDraft = (sIn && typeof sIn.draft === 'string') ? sIn.draft : ''
+      const spaced = (sDraft && !/\s$/.test(sDraft) && !/^\s/.test(text)
+        && !/^[.,;:!?…»)\]}]/.test(text) && !/[\s(«"'\[]$/.test(sDraft)) ? ' ' + text : text
       if (span) {
         if (typeof span.insert === 'function') {
           try {
-            const r = span.insert(text)
+            const r = span.insert(spaced)
             if (r !== false) return true
             return false
           } catch (_) {}
         } else if (typeof span.insertText === 'function') {
           try {
-            const r = span.insertText(text)
+            const r = span.insertText(spaced)
             if (r !== false) return true
             return false
           } catch (_) {}
         } else if (typeof span === 'function') {
           try {
-            const r = span(text)
+            const r = span(spaced)
             if (r !== false) return true
             return false
           } catch (_) {}
@@ -1062,7 +1063,7 @@ window.__ModuleLoader__.load({
       }
       if (typeof actions.insertText === 'function') {
         try {
-          const res = actions.insertText(text, span)
+          const res = actions.insertText(spaced, span)
           if (res !== false) {
             if (target && typeof actions.captureInsertion === 'function') {
               try { target.span = actions.captureInsertion() } catch (_) {}
@@ -1124,14 +1125,7 @@ window.__ModuleLoader__.load({
       } catch (noAudio) { /* no audio — not critical */ }
     }
 
-    // ------------------------------------------------- browser recognition
-    //
-    // A separate leg unlike the others: the browser recognizes speech itself,
-    // nothing is uploaded to the host, no keys are needed, and text appears
-    // word by word while you speak.
-    //
-    // Cost: in Chrome audio goes to Google servers. This provider is never
-    // enabled automatically — only when placed explicitly in a chain.
+    // browser recognition: in-browser, no upload, only enabled when in chain
     function speechRecognitionCtor() {
       if (typeof window === 'undefined') return null
       return window.SpeechRecognition || window.webkitSpeechRecognition || null
@@ -1658,7 +1652,8 @@ voice._chainsPromise = null
         } else {
           rec.sileroSpeech = null
         }
-        if (rec.hadSpeech && raw > 0 && raw < 0.08) {
+        const lowT = Number(voice.settings?.lowVolumeThreshold ?? 0.08)
+        if (rec.hadSpeech && raw > 0 && raw < lowT) {
           rec.lowVolumeUntil = now + 800
           rec.hasLowVolume = true
         }
@@ -3156,6 +3151,7 @@ function cycleDictationMode() {
                 sessionCommands: !!(data && data.modes && data.modes.message && data.modes.message.sessionCommands),
                 noiseSuppression: data && data.noiseSuppression !== false,
                 noiseGateDb: Number(data && data.noiseGateDb !== undefined ? data.noiseGateDb : -45),
+                lowVolumeThreshold: Number(data && data.lowVolumeThreshold !== undefined ? data.lowVolumeThreshold : 0.08),
                 contextGlossary: data && data.contextGlossary !== false,
                 visualizerStyle: (data && data.visualizerStyle) || 'liquid-wave',
                 gatedTurnTaking: data && data.gatedTurnTaking !== false,

package/lib/host-routes.js

@@ -74,6 +74,7 @@ export function registerHostRoutes(ctx, deps) {
         localOnly: cfg.localOnly,
         micDeviceId: cfg.micDeviceId,
         noiseGateDb: cfg.noiseGateDb !== undefined ? cfg.noiseGateDb : -45,
+        lowVolumeThreshold: typeof cfg.lowVolumeThreshold === 'number' ? cfg.lowVolumeThreshold : 0.08,
         historyLimit: cfg.historyLimit,
         voiceCommands: cfg.voiceCommands,
         wakeWord: String(cfg.wakeWord || ''),

package/lib/schema.js

@@ -106,6 +106,8 @@ export const BaseConfig = z.object({
     .description('Microphone device id for recording. Empty means the system default.'),
   noiseGateDb: z.number().default(-45)
     .description('Audio noise gate threshold in dB for client-side recording (-50 to -25 dB, or <= -90 to disable). Silence below this threshold cuts ambient background noise and prevents false VAD triggers.'),
+  lowVolumeThreshold: z.number().min(0).max(1).default(0.08)
+    .description('Recording level (0..1) below which the "Low signal / speak louder" hint appears while dictating. Lower it for quiet microphones; 0 disables the hint.'),
   vadEngine: z.union(['silero', 'amplitude']).default('silero')
     .description('Voice activity detection engine: neural Silero VAD (WASM/ONNX) or classic amplitude threshold.'),
   streamingWs: z.boolean().default(true)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-voice",
-  "version": "0.9.43",
+  "version": "0.9.44",
   "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",