@goodandready/dsh-voice 0.9.36 → 0.9.37

package/lib/index.js

@@ -17,6 +17,7 @@ import { registerConfigRoutes } from './config-routes.js'
 import { registerHostRoutes } from './host-routes.js'
 import { createConfigReader } from './config-compat.js'
 import { createStatsTracker, mergeContextVocabulary } from './stats.js'
+import { buildTranscriptionPrompt } from './transcribe-core.js'
 import { registerTranscribeAudioTool } from './tool.js'
 import { createPolishText } from './polish.js'
 import { buildProviderOrder, sessionCommand } from './transcribe-core.js'
@@ -157,13 +158,17 @@ export function apply(ctx, baseConfig) {
       defaultModels: DEFAULT_MODELS,
     })
 
-    const vocab = cfg.contextGlossary
-      ? mergeContextVocabulary(cfg.vocabulary, contextWords)
-      : (Array.isArray(cfg.vocabulary) ? cfg.vocabulary : [])
+    const { promptText, keywords, vocabulary: vocab } = buildTranscriptionPrompt({
+      vocabulary: cfg.vocabulary,
+      contextWords: cfg.contextGlossary ? contextWords : [],
+      bilingualDeveloperPrompt: cfg.bilingualDeveloperPrompt !== false,
+      bilingualCustomPrompt: cfg.bilingualCustomPrompt || '',
+      language: modeCfg.language,
+    })
 
     const providers = makeProviders(
       { resolveKey, fetchImpl: fetch, cfg, toWav: (b, sig) => toWav16k(b, cfg.ffmpegBin, sig || signal) },
-      { bytes, mime, lang: modeCfg.language, signal, models, vocabulary: vocab },
+      { bytes, mime, lang: modeCfg.language, signal, models, vocabulary: vocab, promptText, keywords },
     )
     return runChain(order, providers, statsTracker.record)
   }

package/lib/providers.js

@@ -87,10 +87,12 @@ function pickModel(models, key) {
 
 export function makeProviders(deps, req) {
   const { resolveKey, fetchImpl, cfg } = deps
-  const { bytes, mime, lang, signal, models } = req
+  const { bytes, mime, lang, signal, models, promptText, keywords } = req
   const vocab = Array.isArray(req.vocabulary)
     ? req.vocabulary.map((w) => String(w || '').trim()).filter(Boolean).join(', ')
     : ''
+  const effectivePrompt = (typeof promptText === 'string' && promptText.trim()) ? promptText.trim() : vocab
+  const effectiveKeywords = (typeof keywords === 'string' && keywords.trim()) ? keywords.trim() : vocab
 
   async function deepgram() {
     const key = await resolveKey(cfg.deepgramKeyEnv)
@@ -99,7 +101,7 @@ export function makeProviders(deps, req) {
     const base = String(cfg.deepgramBaseUrl || 'https://api.deepgram.com').replace(/\/+$/, '')
     const url = `${base}/v1/listen?model=${encodeURIComponent(model)}`
       + (!isAutoLang(lang) ? `&language=${encodeURIComponent(lang)}` : '')
-      + (vocab ? `&keywords=${encodeURIComponent(vocab)}` : '')
+      + (effectiveKeywords ? `&keywords=${encodeURIComponent(effectiveKeywords)}` : '')
       + '&smart_format=true'
     const res = await fetchImpl(url, {
       method: 'POST',
@@ -124,7 +126,7 @@ export function makeProviders(deps, req) {
     form.append('file', new Blob([bytes], { type: mime }), fileName(mime))
     form.append('model', pickModel(models, 'groq'))
     if (!isAutoLang(lang)) form.append('language', lang)
-    if (vocab) form.append('prompt', vocab)
+    if (effectivePrompt) form.append('prompt', effectivePrompt)
     form.append('response_format', 'json')
     const res = await fetchImpl('https://api.groq.com/openai/v1/audio/transcriptions', {
       method: 'POST',
@@ -187,7 +189,7 @@ export function makeProviders(deps, req) {
     // Also override a server started with --translate: this plugin always needs
     // a transcript in the spoken language, never Whisper's English translation.
     form.append('translate', 'false')
-    if (vocab) form.append('prompt', vocab)
+    if (effectivePrompt) form.append('prompt', effectivePrompt)
     form.append('response_format', 'json')
     const res = await fetchImpl(cfg.whisperUrl, { method: 'POST', body: form, signal })
     // whisper.cpp returns 400 with a JSON body on internal failures (e.g. decode
@@ -228,7 +230,7 @@ export function makeProviders(deps, req) {
     if (isOpenAI) {
       form.append('model', pickModel(models, 'sensevoice'))
       if (!isAutoLang(lang)) form.append('language', lang)
-      if (vocab) form.append('prompt', vocab)
+      if (effectivePrompt) form.append('prompt', effectivePrompt)
       form.append('response_format', 'json')
     }
 
@@ -282,7 +284,7 @@ export function makeProviders(deps, req) {
       form.append('file', new Blob([bytes], { type: mime }), fileName(mime))
       form.append('model', model)
       if (lang && lang !== 'auto') form.append('language', lang)
-      if (vocab) form.append('prompt', vocab)
+      if (effectivePrompt) form.append('prompt', effectivePrompt)
       form.append('response_format', 'json')
       const res = await fetchImpl(`${base}/audio/transcriptions`, {
         method: 'POST', headers, body: form, signal,
@@ -303,7 +305,7 @@ export function makeProviders(deps, req) {
         format = 'wav'
       }
       const ask = (spec.prompt || CHAT_AUDIO_PROMPT)
-        + (vocab ? ` Vocabulary hints (spell these correctly): ${vocab}.` : '')
+        + (effectivePrompt ? ` Vocabulary hints (spell these correctly): ${effectivePrompt}.` : '')
         + (lang && lang !== 'auto' ? ` The audio language is ${lang}.` : '')
       const res = await fetchImpl(`${base}/chat/completions`, {
         method: 'POST',

package/lib/schema.js

@@ -112,6 +112,10 @@ export const BaseConfig = z.object({
     .description('Custom words (names, terms) hinted to providers so they recognize them correctly.'),
   contextGlossary: z.boolean().default(true)
     .description('Automatically extract code words and terms from context and hint them to STT providers.'),
+  bilingualDeveloperPrompt: z.boolean().default(true)
+    .description('Inject bilingual developer primer (git, docker, npm, PR, etc.) into the STT prompt so mixed RU/EN technical speech is transcribed accurately.'),
+  bilingualCustomPrompt: z.string().default('')
+    .description('Optional custom terms or prompt primer for the bilingual developer prompt.'),
   noiseSuppression: z.boolean().default(true)
     .description('Enable hardware noise suppression, echo cancellation, and auto gain control.'),
   voiceCommands: z.boolean().default(false)

package/lib/transcribe-core.js

@@ -1,3 +1,4 @@
+import { mergeContextVocabulary } from './stats.js'
 // Pure helpers for the transcribe path: chain selection and session commands.
 
 /**
@@ -47,4 +48,46 @@ export function sessionCommand(text) {
     if (re.test(t)) return cmd
   }
   return null
-}
+}
+
+export const DEFAULT_BILINGUAL_PROMPT = 'TypeScript, Python, JavaScript, Docker, Git commit, PR, push, merge, branch, npm, pnpm, API, JSON, async, await, bug, deploy, Linux, bash. Смешанная речь разработчика: русский язык с английскими терминами, командами и кодом.'
+
+export function buildTranscriptionPrompt({
+  vocabulary = [],
+  contextWords = [],
+  bilingualDeveloperPrompt = true,
+  bilingualCustomPrompt = '',
+  language = 'ru',
+  maxWords = 50,
+} = {}) {
+  const mergedVocab = mergeContextVocabulary(vocabulary, contextWords, maxWords)
+  const isRuOrAuto = !language || language === 'ru' || language === 'auto'
+  let primer = ''
+  if (bilingualDeveloperPrompt && isRuOrAuto) {
+    primer = (typeof bilingualCustomPrompt === 'string' && bilingualCustomPrompt.trim())
+      ? bilingualCustomPrompt.trim()
+      : DEFAULT_BILINGUAL_PROMPT
+  }
+  const vocabStr = mergedVocab.join(', ')
+  let promptText = ''
+  if (primer && vocabStr) {
+    promptText = `${primer} Термины: ${vocabStr}.`
+  } else if (primer) {
+    promptText = primer
+  } else {
+    promptText = vocabStr
+  }
+
+  const keywords = [...mergedVocab]
+  if (bilingualDeveloperPrompt && isRuOrAuto) {
+    for (const w of ['Docker', 'Git', 'commit', 'PR', 'merge', 'branch', 'pnpm', 'npm', 'API', 'JSON', 'TypeScript']) {
+      if (!keywords.includes(w)) keywords.push(w)
+    }
+  }
+
+  return {
+    vocabulary: mergedVocab,
+    promptText: promptText.trim(),
+    keywords: keywords.slice(0, maxWords).join(', '),
+  }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-voice",
-  "version": "0.9.36",
+  "version": "0.9.37",
   "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",