@goodandready/dsh-voice 0.9.34 → 0.9.35

package/lib/client.js

@@ -68,6 +68,9 @@ window.__ModuleLoader__.load({
       'npuProvider': 'Hardware Acceleration (SBC NPU / Host)',
       'npuProviderHint': 'Execution provider for SenseVoice / sherpa-onnx: CPU, Rockchip RK3588 NPU (rknpu), OpenVINO, or CUDA.',
       'npuThreads': 'Threads (CPU/SBC)',
+      'benchmarkTitle': 'Hardware Benchmark (Speed & Latency)',
+      'runBenchmarkBtn': 'Run Benchmark', 'benchmarking': 'Running benchmark…',
+      'benchmarkHint': 'Measures real-time factor (RTF) and inference latency on 1.0s audio for current execution provider.',
       'webgpuStatus': 'Browser WebGPU Status',
       'webgpuAvailable': 'WebGPU Available (hardware acceleration ready)',
       'webgpuNotAvailable': 'WebGPU Not Detected (will fallback to server)',
@@ -291,20 +294,15 @@ window.__ModuleLoader__.load({
       'wizardMicPermissionBad': 'Microphone permission denied',
       'wizardRequestMic': 'Request Microphone Access',
       'wizardMicLabel': 'Microphone input device:',
-      'wizardStartSignalCheck': 'Start Audio Level Test',
-      'wizardStopSignalCheck': 'Stop Audio Test',
+      'wizardStartSignalCheck': 'Start Audio Level Test', 'wizardStopSignalCheck': 'Stop Audio Test',
       'wizardSignalOk': 'Audio signal detected (RMS: {level}%)',
       'wizardSignalZero': 'Zero audio level detected - speak into microphone or select correct device',
       'wizardSignalTesting': 'Listening for audio signal… speak into your microphone',
       'wizardProviderLabel': 'Active STT engine & model:',
       'wizardProcessingArea': 'Processing area:',
-      'wizardAreaBrowser': 'In-browser (WebGPU / WebSpeech)',
-      'wizardAreaServer': 'Local MiniPC server (SenseVoice / Whisper)',
-      'wizardAreaCloud': 'Remote cloud provider',
-      'wizardStartTestRecord': 'Record 3-Second Test Phrase',
-      'wizardRecordingTest': 'Recording test phrase… speak now ({left}s)',
-      'wizardTranscribingTest': 'Transcribing test phrase…',
-      'wizardTestResult': 'STT Test Result:',
+      'wizardAreaBrowser': 'In-browser (WebGPU / WebSpeech)', 'wizardAreaServer': 'Local MiniPC server (SenseVoice / Whisper)', 'wizardAreaCloud': 'Remote cloud provider',
+      'wizardStartTestRecord': 'Record 3-Second Test Phrase', 'wizardRecordingTest': 'Recording test phrase… speak now ({left}s)',
+      'wizardTranscribingTest': 'Transcribing test phrase…', 'wizardTestResult': 'STT Test Result:',
       'wizardTestSuccess': 'STT round-trip verified successfully! Latency: {latency}ms',
       'wizardTestEmpty': 'No speech detected in test recording. Please speak clearly and try again.',
       'wizardTestFailed': 'STT test failed: {error}',
@@ -358,6 +356,9 @@ window.__ModuleLoader__.load({
       'npuProvider': '硬件加速 (SBC NPU / 主机)',
       'npuProviderHint': 'SenseVoice / sherpa-onnx 执行加速引擎：CPU、Rockchip RK3588 NPU (rknpu)、OpenVINO 或 CUDA。',
       'npuThreads': '计算线程 (CPU/SBC)',
+      'benchmarkTitle': '硬件基准测试 (速度与延迟)',
+      'runBenchmarkBtn': '运行基准测试', 'benchmarking': '正在运行测试…',
+      'benchmarkHint': '测试当前执行提供者在 1.0 秒音频下的实时率 (RTF) 与推理延迟。',
       'webgpuStatus': '浏览器 WebGPU 状态',
       'webgpuAvailable': 'WebGPU 可用 (硬件加速就绪)',
       'webgpuNotAvailable': '未检测到 WebGPU (将自动回退到服务端)',
@@ -588,13 +589,9 @@ window.__ModuleLoader__.load({
       'wizardSignalTesting': '正在监听声音信号… 请对着麦克风发声',
       'wizardProviderLabel': '当前语音引擎与模型：',
       'wizardProcessingArea': '处理区域：',
-      'wizardAreaBrowser': '浏览器端处理 (WebGPU / WebSpeech)',
-      'wizardAreaServer': '本地服务器处理 (SenseVoice / Whisper)',
-      'wizardAreaCloud': '远程云端服务',
-      'wizardStartTestRecord': '录制 3 秒测试语句',
-      'wizardRecordingTest': '正在录制测试音频… 请说话 ({left}秒)',
-      'wizardTranscribingTest': '正在请求语音识别…',
-      'wizardTestResult': 'STT 测试结果：',
+      'wizardAreaBrowser': '浏览器端处理 (WebGPU / WebSpeech)', 'wizardAreaServer': '本地服务器处理 (SenseVoice / Whisper)', 'wizardAreaCloud': '远程云端服务',
+      'wizardStartTestRecord': '录制 3 秒测试语句', 'wizardRecordingTest': '正在录制测试音频… 请说话 ({left}秒)',
+      'wizardTranscribingTest': '正在请求语音识别…', 'wizardTestResult': 'STT 测试结果：',
       'wizardTestSuccess': 'STT 链路验证成功！耗时：{latency}毫秒',
       'wizardTestEmpty': '录音中未检测到有效语音，请清晰发音后重试。',
       'wizardTestFailed': 'STT 测试失败：{error}',
@@ -3394,6 +3391,32 @@ function DictationModesSection(props) {
       const gateThresholdPercent = props.gateThresholdPercent
       const onClearKey = props.onClearKey
 
+      const [benchmarking, setBenchmarking] = React.useState(false)
+      const [benchmarkResult, setBenchmarkResult] = React.useState(null)
+      const [benchmarkErr, setBenchmarkErr] = React.useState('')
+
+      const runBenchmark = () => {
+        setBenchmarking(true)
+        setBenchmarkErr('')
+        fetch('/dsh-voice/benchmark', {
+          method: 'POST',
+          headers: {
+            'x-dsh-plugin-update': '1',
+            'Content-Type': 'application/json',
+          },
+        })
+          .then((r) => r.json())
+          .then((d) => {
+            setBenchmarking(false)
+            if (d && d.ok) setBenchmarkResult(d)
+            else setBenchmarkErr((d && d.error && d.error.message) || 'Benchmark failed')
+          })
+          .catch((e) => {
+            setBenchmarking(false)
+            setBenchmarkErr((e && e.message) || 'Network error')
+          })
+      }
+
       const textField = (key, label, hint) => React.createElement('div', { className: 'cb-field' },
         React.createElement('label', null, label),
         React.createElement('input', {
@@ -3564,6 +3587,7 @@ function DictationModesSection(props) {
             },
               React.createElement('option', { value: 'cpu' }, 'CPU (Standard)'),
               React.createElement('option', { value: 'rknn' }, 'Rockchip RK3588 NPU (RKNN)'),
+              React.createElement('option', { value: 'coreml' }, 'Apple Silicon CoreML'),
               React.createElement('option', { value: 'openvino' }, 'Intel OpenVINO'),
               React.createElement('option', { value: 'cuda' }, 'NVIDIA CUDA'),
             ),
@@ -3580,6 +3604,24 @@ function DictationModesSection(props) {
           ),
           React.createElement('span', { className: 'dvs-sub' }, t('npuProviderHint')),
         ),
+        // Hardware Benchmark
+        React.createElement('div', { className: 'cb-field' },
+          React.createElement('div', { className: 'cb-row', style: { alignItems: 'center', justifyContent: 'space-between' } },
+            React.createElement('label', { style: { marginBottom: 0 } }, t('benchmarkTitle')),
+            React.createElement('button', {
+              type: 'button', className: 'cb-btn cb-btn-secondary', style: { fontSize: '12px', padding: '4px 10px' },
+              disabled: benchmarking || !writable, onClick: runBenchmark,
+            }, benchmarking ? ('⏳ ' + t('benchmarking')) : ('⚡ ' + t('runBenchmarkBtn'))),
+          ),
+          benchmarkResult
+            ? React.createElement('div', {
+                style: { marginTop: '8px', padding: '6px 10px', borderRadius: '6px', fontSize: '12px', backgroundColor: 'var(--dsw-alias-fill-quaternary)' }
+              }, benchmarkResult.running
+                ? `⚡ RTF: ${benchmarkResult.rtf}x · ${benchmarkResult.latencyMs}ms (${benchmarkResult.rating}) · ${String(benchmarkResult.provider || 'cpu').toUpperCase()} · Cores: ${benchmarkResult.cpuCores}`
+                : `ℹ ${benchmarkResult.message || 'Offline'} (${benchmarkResult.cpuCores} cores, ${benchmarkResult.platform})`)
+            : (benchmarkErr ? React.createElement('div', { style: { color: 'var(--dsw-alias-state-danger)', fontSize: '12px', marginTop: '4px' } }, benchmarkErr) : null),
+          React.createElement('span', { className: 'dvs-sub' }, t('benchmarkHint')),
+        ),
         // WebGPU Status Indicator
         React.createElement('div', { className: 'cb-field' },
           React.createElement('label', null, t('webgpuStatus')),
@@ -3627,7 +3669,7 @@ function DictationModesSection(props) {
         fetchSensevoice()
       }, [])
 
-      const triggerInstallSensevoice = () => {
+      const triggerInstallSensevoice = (modelType = 'sensevoice') => {
         setInstallingSensevoice(true)
         fetch('/dsh-voice/sensevoice-installer', {
           method: 'POST',
@@ -3635,6 +3677,7 @@ function DictationModesSection(props) {
             'x-dsh-plugin-update': '1',
             'Content-Type': 'application/json',
           },
+          body: JSON.stringify({ model: modelType }),
         })
           .then((r) => r.json())
           .then(() => {
@@ -3933,39 +3976,51 @@ function MacrosSection(props) {
         ),
   )
 }
-    // SenseVoice Local ASR (1-Click Setup)
+    // Local ASR (SenseVoice-Small & Whisper Large-v3-Turbo 1-Click Setup)
     function SensevoiceSection(props) {
       const t = props.t || moduleT
       const sensevoiceState = props.sensevoiceState
       const installingSensevoice = props.installingSensevoice
       const writable = props.writable
       const onInstall = props.onInstall
+      const [modelChoice, setModelChoice] = React.useState('sensevoice')
+
+      const isInstalled = Boolean(sensevoiceState && sensevoiceState.installed)
+      const currentType = (sensevoiceState && sensevoiceState.modelType) || 'sensevoice'
+      const badgeText = isInstalled
+        ? (t('sensevoiceReady') + (currentType === 'whisper-turbo' ? ' (whisper-turbo)' : ' (small-int8)'))
+        : (installingSensevoice ? t('sensevoiceDownloading') : t('sensevoiceNotInstalled'))
 
       return React.createElement('div', { className: 'cb-section-card' },
         React.createElement('div', { className: 'cb-section-title' },
           React.createElement('span', null, '⚡ ' + t('sensevoiceTitle')),
-          sensevoiceState && sensevoiceState.installed
-            ? React.createElement('span', { className: 'cb-badge cb-badge-ok' }, t('sensevoiceReady') + ' (small-int8)')
-            : React.createElement('span', { className: 'cb-badge cb-badge-warn' },
-                installingSensevoice ? t('sensevoiceDownloading') : t('sensevoiceNotInstalled')
-              ),
+          React.createElement('span', { className: 'cb-badge ' + (isInstalled ? 'cb-badge-ok' : 'cb-badge-warn') }, badgeText),
         ),
         React.createElement('div', { className: 'cb-section-desc' }, t('sensevoiceDesc')),
         React.createElement('div', { className: 'dvo-installer-box' },
           React.createElement('div', { className: 'cb-row', style: { alignItems: 'center', justifyContent: 'space-between', gap: '12px' } },
-            React.createElement('div', { style: { fontSize: '12px', color: 'var(--dsw-alias-content-secondary)' } },
-              sensevoiceState && sensevoiceState.path
-                ? ('~/' + sensevoiceState.path.replace(/^.*[\\/](\.dsh[\\/]models[\\/].*)$/, '$1'))
-                : '~/.dsh/models/sensevoice'
+            React.createElement('select', {
+              value: modelChoice,
+              disabled: installingSensevoice || !writable,
+              onChange: (e) => setModelChoice(e.target.value),
+              style: { flex: 1, fontSize: '12px', minWidth: '180px' },
+            },
+              React.createElement('option', { value: 'sensevoice' }, 'SenseVoice-Small (150MB, ~50ms)'),
+              React.createElement('option', { value: 'whisper-turbo' }, 'Whisper Large-v3-Turbo (800MB, High-Accuracy)'),
             ),
             React.createElement('button', {
               type: 'button',
               className: 'cb-btn cb-btn-primary',
-              style: { fontSize: '12px', padding: '6px 12px' },
+              style: { fontSize: '12px', padding: '6px 12px', whiteSpace: 'nowrap' },
               disabled: installingSensevoice || !writable,
-              onClick: onInstall,
+              onClick: () => { if (typeof onInstall === 'function') onInstall(modelChoice) },
             }, installingSensevoice ? ('⏳ ' + t('installingSensevoice')) : ('📥 ' + t('installSensevoice'))),
           ),
+          React.createElement('div', { style: { fontSize: '11px', marginTop: '6px', color: 'var(--dsw-alias-content-secondary)' } },
+            sensevoiceState && sensevoiceState.path
+              ? ('~/' + sensevoiceState.path.replace(/^.*[\\/](\.dsh[\\/]models[\\/].*)$/, '$1'))
+              : ('~/.dsh/models/' + modelChoice)
+          ),
           installingSensevoice
             ? React.createElement('div', { style: { marginTop: '8px' } },
                 React.createElement('div', { className: 'dvo-progress-bar' },

package/lib/host-routes.js

@@ -1,7 +1,9 @@
 // HTTP host routes registered on ctx.webServer.
+import os from 'node:os'
 import { readBody, isLoopback, isTrustedCaller } from './http-util.js'
 import { resolveSensevoiceProvider } from './sensevoice-installer.js'
 import { applyVoiceMacros } from './macros.js'
+import { createSyntheticWav } from './wav.js'
 
 function writeJson(res, status, obj) {
   res.writeHead(status, { 'content-type': 'application/json' })
@@ -147,6 +149,78 @@ export function registerHostRoutes(ctx, deps) {
     },
   }), 'dsh-voice: /diagnostics/export route')
 
+  ctx.effect(() => ctx.webServer.register({
+    kind: 'exact',
+    path: '/dsh-voice/benchmark',
+    handler: async (req, res) => {
+      if (req.method !== 'POST') { writeJson(res, 405, { ok: false, error: { code: 'method', message: 'POST only' } }); return }
+      if (!isTrustedCaller(req)) { writeJson(res, 403, { ok: false, error: { code: 'forbidden', message: 'forbidden source origin' } }); return }
+
+      const cfg = live()
+      const effectiveProvider = resolveSensevoiceProvider(cfg)
+      const cpus = os.cpus() || []
+      const numCores = cpus.length || 1
+      const cpuModel = (cpus[0] && cpus[0].model) || 'unknown'
+      const platformInfo = `${process.platform}-${process.arch}`
+      const isAlive = await sensevoiceAlive()
+
+      if (!isAlive) {
+        writeJson(res, 200, {
+          ok: true,
+          running: false,
+          provider: effectiveProvider,
+          cpuCores: numCores,
+          cpuModel,
+          platform: platformInfo,
+          message: 'Local ASR server is not running. Hardware detected.',
+        })
+        return
+      }
+
+      const sampleSec = 1.0
+      const wavBuf = createSyntheticWav(sampleSec, 16000)
+      const tStart = Date.now()
+      let latencyMs = 0
+      let recognizedText = ''
+
+      try {
+        const url = String(cfg.sensevoiceUrl || 'http://127.0.0.1:6006/api/v1/asr').trim()
+        const form = new FormData()
+        form.append('file', new Blob([wavBuf], { type: 'audio/wav' }), 'benchmark.wav')
+        const r = await fetch(url, { method: 'POST', body: form, signal: AbortSignal.timeout(10000) })
+        latencyMs = Date.now() - tStart
+        if (r.ok) {
+          try {
+            const data = await r.json()
+            recognizedText = String(data?.text || data?.transcript || '').trim()
+          } catch { /* json parse error */ }
+        }
+      } catch {
+        latencyMs = Date.now() - tStart
+      }
+
+      const rtf = Math.round((latencyMs / (sampleSec * 1000)) * 1000) / 1000
+      let rating = 'Fast'
+      if (rtf < 0.15) rating = 'Ultra-fast'
+      else if (rtf < 0.45) rating = 'Fast'
+      else if (rtf <= 1.0) rating = 'Real-time'
+      else rating = 'Slow'
+
+      writeJson(res, 200, {
+        ok: true,
+        running: true,
+        provider: effectiveProvider,
+        latencyMs,
+        rtf,
+        sampleSec,
+        rating,
+        cpuCores: numCores,
+        cpuModel,
+        platform: platformInfo,
+      })
+    },
+  }), 'dsh-voice: /benchmark route')
+
   ctx.effect(() => ctx.webServer.register({
     kind: 'exact',
     path: '/dsh-voice/transcribe',

package/lib/schema.js

@@ -139,7 +139,7 @@ export const BaseConfig = z.object({
   sensevoiceAutostart: z.boolean().default(false)
     .description('Start the local SenseVoice / sherpa-onnx server process automatically on startup.'),
   sensevoiceProvider: z.string().default('cpu')
-    .description('Execution provider for sherpa-onnx: "cpu", "rknn" (Rockchip RK3588/RK3568 NPU on ARM SBC with RKNN models), "openvino", "cuda".'),
+    .description('Execution provider for sherpa-onnx: "cpu", "coreml" (Apple Silicon), "openvino", "cuda", "rknn" (Rockchip RK3588/RK3568 NPU on ARM SBC with RKNN models).'),
   sensevoiceThreads: z.number().default(4)
     .description('Execution thread count for local SenseVoice / sherpa-onnx on CPU/SBC.'),
   webgpuModel: z.string().default('onnx-community/whisper-tiny')

package/lib/sensevoice-installer.js

@@ -1,5 +1,5 @@
-// lib/sensevoice-installer.js — 1-Click installer for SenseVoice-Small / sherpa-onnx
-// Downloads pre-quantized ONNX model and configures autostart on the host.
+// lib/sensevoice-installer.js — 1-Click installer for SenseVoice-Small & Whisper Large-v3-Turbo
+// Downloads pre-quantized ONNX models and configures autostart on the host via sherpa-onnx.
 
 import { createWriteStream, existsSync, readdirSync } from 'node:fs'
 import { mkdir, unlink } from 'node:fs/promises'
@@ -8,25 +8,63 @@ import path from 'node:path'
 import { spawn } from 'node:child_process'
 import { Readable, Transform } from 'node:stream'
 import { pipeline } from 'node:stream/promises'
-import { writeJson, isTrustedCaller } from './http-util.js'
+import { writeJson, readBody, isTrustedCaller } from './http-util.js'
 
 export const SENSEVOICE_MODEL_URL =
   'https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17.tar.bz2'
 export const SENSEVOICE_MODEL_MIRROR =
   'https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17.tar.bz2'
 
+export const WHISPER_TURBO_MODEL_URL =
+  'https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-turbo.tar.bz2'
+export const WHISPER_TURBO_MODEL_MIRROR =
+  'https://huggingface.co/csukuangfj/sherpa-onnx-whisper-turbo/resolve/main/sherpa-onnx-whisper-turbo.tar.bz2'
+
 let installerState = {
   installing: false,
   progress: 0,
   step: 'idle',
   error: null,
   modelPath: '',
+  modelType: 'sensevoice',
 }
 
-export function getDefaultModelDir() {
+export function getDefaultModelDir(type = 'sensevoice') {
+  const norm = String(type || '').toLowerCase()
+  if (norm.includes('whisper') || norm.includes('turbo')) {
+    return path.join(homedir(), '.dsh', 'models', 'whisper-turbo')
+  }
   return path.join(homedir(), '.dsh', 'models', 'sensevoice')
 }
 
+export function findWhisperFiles(dir) {
+  if (!dir || !existsSync(dir)) return null
+  try {
+    const entries = readdirSync(dir, { recursive: true, withFileTypes: true })
+    let encoder = null
+    let decoder = null
+    let tokens = null
+    for (const entry of entries) {
+      if (!entry.isFile()) continue
+      const name = entry.name.toLowerCase()
+      const parent = entry.parentPath || (entry.path ? entry.path : dir)
+      const fullPath = path.join(parent, entry.name)
+
+      if (name.includes('encoder') && name.endsWith('.onnx')) {
+        if (!encoder || name.includes('int8')) encoder = fullPath
+      } else if (name.includes('decoder') && name.endsWith('.onnx')) {
+        if (!decoder || name.includes('int8')) decoder = fullPath
+      } else if (name.endsWith('tokens.txt') || (name.includes('tokens') && name.endsWith('.txt'))) {
+        tokens = fullPath
+      }
+    }
+    if (encoder && decoder) {
+      return { encoder, decoder, tokens }
+    }
+  } catch { /* directory read failed */ }
+  return null
+}
+
 export function findModelFile(dir) {
   if (!dir || !existsSync(dir)) return null
   try {
@@ -37,23 +75,26 @@ export function findModelFile(dir) {
         return path.join(parent, entry.name)
       }
     }
+    const whisper = findWhisperFiles(dir)
+    if (whisper) return whisper.encoder
   } catch { /* directory read failed */ }
   return null
 }
 
 export function redactModelPath(filePath) {
-  if (!filePath) return ""
-  const str = String(filePath).replace(/\\/g, "/")
-  const idx = str.indexOf(".dsh/models/")
-  if (idx !== -1) return "~/" + str.slice(idx)
+  if (!filePath) return ''
+  const str = String(filePath).replace(/\\/g, '/')
+  const idx = str.indexOf('.dsh/models/')
+  if (idx !== -1) return '~/' + str.slice(idx)
   return path.basename(str)
 }
+
 export function findTokensFile(dir) {
   if (!dir || !existsSync(dir)) return null
   try {
     const entries = readdirSync(dir, { recursive: true, withFileTypes: true })
     for (const entry of entries) {
-      if (entry.isFile() && entry.name === 'tokens.txt') {
+      if (entry.isFile() && (entry.name === 'tokens.txt' || entry.name === 'turbo-tokens.txt')) {
         const parent = entry.parentPath || (entry.path ? entry.path : dir)
         return path.join(parent, entry.name)
       }
@@ -63,21 +104,49 @@ export function findTokensFile(dir) {
 }
 
 export async function getSensevoiceStatus(cfg, isAliveFn) {
-  const modelDir = cfg?.sensevoiceModel ? path.dirname(cfg.sensevoiceModel) : getDefaultModelDir()
-  const modelFile = cfg?.sensevoiceModel && existsSync(cfg.sensevoiceModel)
-    ? cfg.sensevoiceModel
-    : findModelFile(getDefaultModelDir())
-  const tokensFile = modelFile ? findTokensFile(path.dirname(modelFile)) : null
+  const customModel = cfg?.sensevoiceModel
+  const customExists = Boolean(customModel && existsSync(customModel))
+  const isWhisperCustom = customExists && (
+    customModel.toLowerCase().includes('whisper') ||
+    customModel.toLowerCase().includes('turbo') ||
+    Boolean(findWhisperFiles(path.dirname(customModel)))
+  )
+
+  const svDir = getDefaultModelDir('sensevoice')
+  const svFile = findModelFile(svDir)
+  const turboDir = getDefaultModelDir('whisper-turbo')
+  const turboFiles = findWhisperFiles(turboDir)
+
+  let activeModelFile = null
+  let activeTokensFile = null
+  let activeModelType = 'sensevoice'
+
+  if (customExists) {
+    activeModelFile = customModel
+    activeTokensFile = findTokensFile(path.dirname(customModel))
+    activeModelType = isWhisperCustom ? 'whisper-turbo' : 'sensevoice'
+  } else if (turboFiles) {
+    activeModelFile = turboFiles.encoder
+    activeTokensFile = turboFiles.tokens
+    activeModelType = 'whisper-turbo'
+  } else if (svFile) {
+    activeModelFile = svFile
+    activeTokensFile = findTokensFile(path.dirname(svFile))
+    activeModelType = 'sensevoice'
+  }
 
   const running = typeof isAliveFn === 'function' ? await isAliveFn() : false
-  const installed = Boolean(modelFile && existsSync(modelFile))
+  const installed = Boolean(activeModelFile && existsSync(activeModelFile))
 
   return {
     installed,
     running,
-    modelPath: redactModelPath(modelFile),
-    tokensPath: redactModelPath(tokensFile),
-    path: installed ? '~/.dsh/models/sensevoice' : '',
+    modelType: activeModelType,
+    modelPath: redactModelPath(activeModelFile),
+    tokensPath: redactModelPath(activeTokensFile),
+    path: installed ? `~/.dsh/models/${activeModelType}` : '',
+    sensevoiceInstalled: Boolean(svFile),
+    whisperTurboInstalled: Boolean(turboFiles),
     installing: installerState.installing,
     progress: installerState.progress,
     step: installerState.step,
@@ -118,21 +187,28 @@ export function extractTarBz2(archivePath, targetDir) {
   })
 }
 
-export async function startSensevoiceInstall(ctx, cfg, updateConfigFn, startSensevoiceFn) {
+export async function startSensevoiceInstall(ctx, cfg, updateConfigFn, startSensevoiceFn, options = {}) {
   if (installerState.installing) {
-    throw new Error('SenseVoice installation is already in progress')
+    throw new Error('ASR model installation is already in progress')
   }
 
+  const targetModel = String(options?.model || 'sensevoice').toLowerCase().includes('whisper')
+    ? 'whisper-turbo'
+    : 'sensevoice'
+
   installerState = {
     installing: true,
     progress: 5,
     step: 'preparing',
     error: null,
     modelPath: '',
+    modelType: targetModel,
   }
 
-  const modelDir = getDefaultModelDir()
-  const archivePath = path.join(modelDir, 'sensevoice.tar.bz2')
+  const modelDir = getDefaultModelDir(targetModel)
+  const archivePath = path.join(modelDir, `${targetModel}.tar.bz2`)
+  const primaryUrl = targetModel === 'whisper-turbo' ? WHISPER_TURBO_MODEL_URL : SENSEVOICE_MODEL_URL
+  const mirrorUrl = targetModel === 'whisper-turbo' ? WHISPER_TURBO_MODEL_MIRROR : SENSEVOICE_MODEL_MIRROR
 
   try {
     await mkdir(modelDir, { recursive: true })
@@ -141,30 +217,36 @@ export async function startSensevoiceInstall(ctx, cfg, updateConfigFn, startSens
 
     let downloadOk = false
     try {
-      await downloadArchive(SENSEVOICE_MODEL_URL, archivePath, (pct, status) => {
+      await downloadArchive(primaryUrl, archivePath, (pct, status) => {
         installerState.progress = pct
         installerState.step = status
       })
       downloadOk = true
-    } catch (e1) {
-      // Fallback to mirror
+    } catch {
       installerState.step = 'downloading (mirror)'
-      await downloadArchive(SENSEVOICE_MODEL_MIRROR, archivePath, (pct, status) => {
+      await downloadArchive(mirrorUrl, archivePath, (pct, status) => {
         installerState.progress = pct
         installerState.step = status
       })
       downloadOk = true
     }
 
-    if (!downloadOk) throw new Error('Could not download SenseVoice model from primary or mirror URL')
+    if (!downloadOk) throw new Error(`Could not download ${targetModel} model from primary or mirror URL`)
 
     installerState.step = 'extracting'
     installerState.progress = 95
     await extractTarBz2(archivePath, modelDir)
     await unlink(archivePath).catch(() => {})
 
-    const modelFile = findModelFile(modelDir)
-    if (!modelFile) throw new Error('model.int8.onnx not found after extraction')
+    let modelFile = null
+    if (targetModel === 'whisper-turbo') {
+      const files = findWhisperFiles(modelDir)
+      if (!files) throw new Error('whisper-turbo encoder/decoder not found after extraction')
+      modelFile = files.encoder
+    } else {
+      modelFile = findModelFile(modelDir)
+      if (!modelFile) throw new Error('model.int8.onnx not found after extraction')
+    }
 
     installerState.modelPath = modelFile
     installerState.progress = 100
@@ -182,7 +264,7 @@ export async function startSensevoiceInstall(ctx, cfg, updateConfigFn, startSens
       await startSensevoiceFn()
     }
 
-    return { ok: true, modelPath: modelFile }
+    return { ok: true, modelType: targetModel, modelPath: modelFile }
   } catch (err) {
     await unlink(archivePath).catch(() => {})
     installerState.installing = false
@@ -217,12 +299,18 @@ export function registerSensevoiceInstaller(ctx, { liveConfig, isAlive, updateCo
           return
         }
 
+        let bodyObj = {}
+        try {
+          const raw = await readBody(req, 16384)
+          bodyObj = JSON.parse(raw.toString('utf8') || '{}')
+        } catch { bodyObj = {} }
+
         // Start background install
-        startSensevoiceInstall(ctx, liveConfig(), updateConfig, startSensevoice).catch((err) => {
-          ctx.logger?.warn?.(`SenseVoice 1-click install error: ${String(err)}`)
+        startSensevoiceInstall(ctx, liveConfig(), updateConfig, startSensevoice, bodyObj).catch((err) => {
+          ctx.logger?.warn?.(`ASR 1-click install error: ${String(err)}`)
         })
 
-        writeJson(res, 202, { ok: true, message: 'SenseVoice installation started', state: safeState })
+        writeJson(res, 202, { ok: true, message: 'ASR installation started', state: safeState })
         return
       }
 
@@ -245,50 +333,59 @@ export function resolveSensevoiceProvider(cfg) {
       if (existsSync(rknnCandidate) || existsSync(defaultRknn)) {
         return 'rknn'
       }
-      // Guard: ordinary ONNX model cannot be loaded by sherpa-onnx RKNN runtime (#221)
       return 'cpu'
     }
     if (ext === '.rknn') {
       if (modelPath && existsSync(modelPath)) {
         return 'rknn'
       }
-      // Guard: missing .rknn model file cannot be claimed as running accelerator (#168)
       return 'cpu'
     }
     return 'cpu'
   }
 
+  if (['cpu', 'openvino', 'coreml', 'cuda'].includes(normProv)) {
+    return normProv
+  }
+
   return normProv || 'cpu'
 }
 
 export function buildSensevoiceArgs(cfg, port) {
   let modelPath = String(cfg?.sensevoiceModel || '')
-  const modelDir = modelPath ? path.dirname(modelPath) : getDefaultModelDir()
-  const tokensPath = (cfg?.sensevoiceTokens && existsSync(cfg.sensevoiceTokens))
-    ? cfg.sensevoiceTokens
-    : findTokensFile(modelDir)
-  const tokensArg = tokensPath ? ` --tokens=${JSON.stringify(tokensPath)}` : ''
+  const modelDir = modelPath ? (existsSync(modelPath) ? path.dirname(modelPath) : modelPath) : getDefaultModelDir()
   const effectiveProv = resolveSensevoiceProvider(cfg)
+  const provider = (effectiveProv && effectiveProv !== 'cpu') ? ` --provider=${effectiveProv}` : ''
+  const threads = Number(cfg?.sensevoiceThreads) || 4
+
+  const whisperFiles = findWhisperFiles(modelDir) || (modelPath ? findWhisperFiles(path.dirname(modelPath)) : null)
+  const isWhisper = Boolean(
+    whisperFiles ||
+    modelPath.toLowerCase().includes('whisper') ||
+    modelPath.toLowerCase().includes('turbo')
+  )
+
+  if (isWhisper && whisperFiles) {
+    const enc = JSON.stringify(whisperFiles.encoder)
+    const dec = JSON.stringify(whisperFiles.decoder)
+    const tok = whisperFiles.tokens ? ` --tokens=${JSON.stringify(whisperFiles.tokens)}` : ''
+    return ` --whisper-encoder=${enc} --whisper-decoder=${dec}${tok} --port=${port} --num-threads=${threads}${provider}`
+  }
 
   if (effectiveProv === 'rknn') {
     const ext = path.extname(modelPath).toLowerCase()
     if (ext === '.onnx') {
       const rknnCandidate = path.join(modelDir, path.basename(modelPath, ext) + '.rknn')
       const defaultRknn = path.join(modelDir, 'model.rknn')
-      if (existsSync(rknnCandidate)) {
-        modelPath = rknnCandidate
-      } else if (existsSync(defaultRknn)) {
-        modelPath = defaultRknn
-      }
+      if (existsSync(rknnCandidate)) modelPath = rknnCandidate
+      else if (existsSync(defaultRknn)) modelPath = defaultRknn
     }
   }
 
-  const provider = (effectiveProv && effectiveProv !== 'cpu')
-    ? ` --provider=${effectiveProv}`
-    : ''
-  const threads = Number(cfg?.sensevoiceThreads) || 4
-  return ` --sense-voice-model=${JSON.stringify(modelPath)}`
-    + tokensArg
-    + ` --port=${port} --num-threads=${threads}`
-    + provider
+  const tokensPath = (cfg?.sensevoiceTokens && existsSync(cfg.sensevoiceTokens))
+    ? cfg.sensevoiceTokens
+    : findTokensFile(modelDir)
+  const tokensArg = tokensPath ? ` --tokens=${JSON.stringify(tokensPath)}` : ''
+
+  return ` --sense-voice-model=${JSON.stringify(modelPath)}${tokensArg} --port=${port} --num-threads=${threads}${provider}`
 }

package/lib/wav.js

@@ -70,3 +70,30 @@ export function toWav16k(bytes, ffmpegBin = 'ffmpeg', signal = null) {
     proc.stdin.end(Buffer.from(bytes))
   })
 }
+
+/**
+ * Generate a synthetic PCM 16-bit mono WAV buffer (silence).
+ * Used for express hardware benchmarking and latency profiling without external files.
+ * @param {number} durationSeconds
+ * @param {number} sampleRate
+ * @returns {Buffer}
+ */
+export function createSyntheticWav(durationSeconds = 1.0, sampleRate = 16000) {
+  const numSamples = Math.max(1, Math.floor(durationSeconds * sampleRate))
+  const dataSize = numSamples * 2
+  const buf = Buffer.alloc(44 + dataSize)
+  buf.write('RIFF', 0)
+  buf.writeUInt32LE(36 + dataSize, 4)
+  buf.write('WAVE', 8)
+  buf.write('fmt ', 12)
+  buf.writeUInt32LE(16, 16)
+  buf.writeUInt16LE(1, 20) // PCM format
+  buf.writeUInt16LE(1, 22) // Mono
+  buf.writeUInt32LE(sampleRate, 24)
+  buf.writeUInt32LE(sampleRate * 2, 28) // Byte rate
+  buf.writeUInt16LE(2, 32) // Block align
+  buf.writeUInt16LE(16, 34) // 16-bit
+  buf.write('data', 36)
+  buf.writeUInt32LE(dataSize, 40)
+  return buf
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-voice",
-  "version": "0.9.34",
+  "version": "0.9.35",
   "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",