@goodandready/dsh-voice 0.9.28 → 0.9.29

package/lib/client.js

@@ -1080,39 +1080,73 @@ window.__ModuleLoader__.load({
     // Fetch the mode chain from the host once. Only needed to decide whether
 
     voice._core = { tidyPhrase, applyVoiceCommands, VOICE_COMMANDS, undoLastInsert, insertHistory, insertDraftText, appendDraft, resolveBrowserRecognitionLang }
-// lib/client-src/35-webgpu.js — Client-side WebGPU Whisper offline recognition
-// Detects WebGPU availability in browser and provides offline transcription with graceful fallback.
+// lib/client-src/35-webgpu.js — Client-side WebGPU/WASM Whisper offline recognition
+// Detects WebGPU and WebAssembly availability in browser and provides offline transcription with graceful WASM fallback.
 
 function isWebGpuSupported() {
   if (typeof navigator === 'undefined' || !navigator.gpu) return false
   return true
 }
 
+function isWasmSupported() {
+  return typeof WebAssembly !== 'undefined'
+}
+
+function isOfflineSupported() {
+  return isWebGpuSupported() || isWasmSupported()
+}
+
 const _webGpuPipelines = new Map()
 
-function loadWebGpuWhisper(modelName) {
+function loadWebGpuWhisper(modelName, options) {
   const targetModel = modelName || (voice.settings && voice.settings.webgpuModel) || 'onnx-community/whisper-tiny'
-  if (_webGpuPipelines.has(targetModel)) {
-    return _webGpuPipelines.get(targetModel)
+  const forceDevice = options?.device
+  const key = `${targetModel}:${forceDevice || 'auto'}`
+  if (_webGpuPipelines.has(key)) {
+    return _webGpuPipelines.get(key)
   }
   const promise = (async () => {
-    if (!isWebGpuSupported()) {
-      throw new Error('WebGPU is not supported in this browser environment')
+    if (!isOfflineSupported()) {
+      throw new Error('Neither WebGPU nor WebAssembly is supported in this browser environment')
     }
     const mod = typeof globalThis !== 'undefined' && globalThis.__mockTransformers
       ? globalThis.__mockTransformers
       : (await import('https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.3.3'))
     const pipeline = mod.pipeline || (mod.default && mod.default.pipeline)
     if (!pipeline) throw new Error('Transformers pipeline not found in loaded bundle')
-    return pipeline('automatic-speech-recognition', targetModel, {
-      device: 'webgpu',
-      dtype: 'fp32',
-    })
+
+    // Prefer WebGPU if available and not forced to wasm; fall back to WASM on error or when WebGPU is absent
+    const canGpu = forceDevice === 'webgpu' || (!forceDevice && isWebGpuSupported())
+    if (canGpu) {
+      try {
+        return await pipeline('automatic-speech-recognition', targetModel, {
+          device: 'webgpu',
+          dtype: 'fp32',
+        })
+      } catch (gpuErr) {
+        if (isWasmSupported()) {
+          return await pipeline('automatic-speech-recognition', targetModel, {
+            device: 'wasm',
+            dtype: 'q8',
+          })
+        }
+        throw gpuErr
+      }
+    }
+
+    if (isWasmSupported()) {
+      return await pipeline('automatic-speech-recognition', targetModel, {
+        device: 'wasm',
+        dtype: 'q8',
+      })
+    }
+
+    throw new Error('Offline speech recognition failed: no WebGPU or WebAssembly available')
   })().catch((err) => {
-    _webGpuPipelines.delete(targetModel)
+    _webGpuPipelines.delete(key)
     throw err
   })
-  _webGpuPipelines.set(targetModel, promise)
+  _webGpuPipelines.set(key, promise)
   return promise
 }
 
@@ -1131,10 +1165,10 @@ async function audioBlobToFloat32Array(blob) {
 }
 
 async function transcribeWithWebGpu(blob, mime, options) {
-  if (!isWebGpuSupported()) {
-    throw new Error('WebGPU unavailable in browser')
+  if (!isOfflineSupported()) {
+    throw new Error('Offline recognition (WebGPU/WASM) unavailable in browser')
   }
-  const transcriber = await loadWebGpuWhisper(options && options.model)
+  const transcriber = await loadWebGpuWhisper(options && options.model, options)
   const audioData = await audioBlobToFloat32Array(blob)
   const generateArgs = {}
   if (options && options.language && options.language !== 'auto') {
@@ -1146,7 +1180,9 @@ async function transcribeWithWebGpu(blob, mime, options) {
 }
 
 voice.webGpu = {
-  isSupported: isWebGpuSupported,
+  isSupported: isOfflineSupported,
+  isWebGpuSupported,
+  isWasmSupported,
   transcribe: transcribeWithWebGpu,
   load: loadWebGpuWhisper,
   clearCache: () => { _webGpuPipelines.clear() },
@@ -3997,6 +4033,9 @@ voice._chainsPromise = null
       stopCurrent,
       submitPending,
       loadWebGpuWhisper,
+      isWebGpuSupported,
+      isWasmSupported,
+      isOfflineSupported,
       waitStop,
       currentLevel,
       accentColor,

package/lib/sensevoice-installer.js

@@ -249,8 +249,13 @@ export function resolveSensevoiceProvider(cfg) {
       return 'cpu'
     }
     if (ext === '.rknn') {
-      return 'rknn'
+      if (modelPath && existsSync(modelPath)) {
+        return 'rknn'
+      }
+      // Guard: missing .rknn model file cannot be claimed as running accelerator (#168)
+      return 'cpu'
     }
+    return 'cpu'
   }
 
   return normProv || 'cpu'
@@ -263,10 +268,9 @@ export function buildSensevoiceArgs(cfg, port) {
     ? cfg.sensevoiceTokens
     : findTokensFile(modelDir)
   const tokensArg = tokensPath ? ` --tokens=${JSON.stringify(tokensPath)}` : ''
-  const rawProv = String(cfg?.sensevoiceProvider || '').trim().toLowerCase()
-  let normProv = rawProv === 'rknpu' ? 'rknn' : rawProv
+  const effectiveProv = resolveSensevoiceProvider(cfg)
 
-  if (normProv === 'rknn') {
+  if (effectiveProv === 'rknn') {
     const ext = path.extname(modelPath).toLowerCase()
     if (ext === '.onnx') {
       const rknnCandidate = path.join(modelDir, path.basename(modelPath, ext) + '.rknn')
@@ -275,15 +279,12 @@ export function buildSensevoiceArgs(cfg, port) {
         modelPath = rknnCandidate
       } else if (existsSync(defaultRknn)) {
         modelPath = defaultRknn
-      } else {
-        // Guard: ordinary ONNX model cannot be loaded by sherpa-onnx RKNN runtime (#221)
-        normProv = 'cpu'
       }
     }
   }
 
-  const provider = (normProv && normProv !== 'cpu')
-    ? ` --provider=${normProv}`
+  const provider = (effectiveProv && effectiveProv !== 'cpu')
+    ? ` --provider=${effectiveProv}`
     : ''
   const threads = Number(cfg?.sensevoiceThreads) || 4
   return ` --sense-voice-model=${JSON.stringify(modelPath)}`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-voice",
-  "version": "0.9.28",
+  "version": "0.9.29",
   "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",