dsh-voice-mode 0.2.2 → 0.3.0

package/lib/sense-worker.mjs

@@ -0,0 +1,128 @@
+// src/sense-worker.ts
+import { parentPort, workerData } from "node:worker_threads";
+function createSenseWorkerClient(worker) {
+  let counter = 0;
+  const pending = /* @__PURE__ */ new Map();
+  let dead = false;
+  const deathFns = /* @__PURE__ */ new Set();
+  const die = () => {
+    if (dead) return;
+    dead = true;
+    for (const fn of deathFns) {
+      try {
+        fn();
+      } catch {
+      }
+    }
+  };
+  worker.on?.("message", (msg) => {
+    const p = pending.get(msg?.id);
+    if (!p) return;
+    pending.delete(msg.id);
+    if (!msg.ok) {
+      p.resolve(null);
+      return;
+    }
+    p.resolve(p.op === "create" ? true : msg.text ?? "");
+  });
+  worker.on?.("error", (e) => {
+    die();
+    const err = new Error("sense worker error: " + String(e?.message ?? e));
+    for (const [, p] of pending) p.reject(err);
+    pending.clear();
+  });
+  worker.on?.("exit", () => {
+    die();
+    const err = new Error("sense worker exited");
+    for (const [, p] of pending) p.reject(err);
+    pending.clear();
+  });
+  const request = (op, samples) => {
+    if (dead) return Promise.reject(new Error("sense worker dead"));
+    const id = counter++;
+    return new Promise((resolve, reject) => {
+      pending.set(id, { op, resolve, reject });
+      const msg = { id, op };
+      if (samples) msg.samples = samples;
+      try {
+        worker.postMessage(msg);
+      } catch (e) {
+        pending.delete(id);
+        reject(e instanceof Error ? e : new Error(String(e)));
+      }
+    });
+  };
+  return {
+    request,
+    onDeath(fn) {
+      deathFns.add(fn);
+    },
+    terminate: async () => {
+      dead = true;
+      const err = new Error("sense worker terminated");
+      for (const [, p] of pending) p.reject(err);
+      pending.clear();
+      try {
+        await worker.terminate?.();
+      } catch {
+      }
+    }
+  };
+}
+function startSenseWorker(data) {
+  const port = parentPort;
+  if (!port) return;
+  let recognizer = null;
+  let sherpa = null;
+  port.on("message", async (msg) => {
+    try {
+      if (msg.op === "create" || msg.op === "decode") {
+        if (!sherpa) {
+          sherpa = await import(data.sherpaModule);
+        }
+        if (!recognizer) {
+          recognizer = sherpa.createOfflineRecognizer({
+            featConfig: { sampleRate: 16e3, featureDim: 80 },
+            modelConfig: {
+              senseVoice: {
+                model: data.modelDir + "/model.int8.onnx",
+                language: "auto",
+                useInverseTextNormalization: 1
+              },
+              tokens: data.modelDir + "/tokens.txt",
+              provider: "cpu",
+              debug: 0
+            }
+          });
+        }
+        if (msg.op === "decode" && msg.samples) {
+          const stream = recognizer.createStream();
+          try {
+            stream.acceptWaveform(16e3, msg.samples);
+            recognizer.decode(stream);
+            const text = recognizer.getResult(stream).text.trim();
+            port.postMessage({ id: msg.id, ok: true, text });
+          } finally {
+            try {
+              stream.free();
+            } catch {
+            }
+          }
+          return;
+        }
+        port.postMessage({ id: msg.id, ok: true, text: "" });
+        return;
+      }
+      port.postMessage({ id: msg.id, ok: false, error: "unknown op: " + msg.op });
+    } catch (e) {
+      port.postMessage({ id: msg.id, ok: false, error: String(e) });
+    }
+  });
+}
+if (parentPort) {
+  startSenseWorker(workerData);
+}
+export {
+  createSenseWorkerClient,
+  startSenseWorker
+};

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-voice-mode",
   "description": "Full-duplex voice mode for DeepSeek Harness: zipformer2 streaming ASR → editable draft, Edge TTS sentence-by-sentence read-aloud with live captions, true barge-in — on-device ASR, no API key. · DSH 语音双工对话：流式识别入草稿、按句朗读+实时字幕、开口即打断，识别本地推理、无需 API Key",
-  "version": "0.2.2",
+  "version": "0.3.0",
   "type": "module",
   "main": "lib/index.js",
   "repository": {
@@ -24,11 +24,14 @@
   "files": [
     "lib/index.js",
     "lib/client.js",
+    "lib/sense-worker.mjs",
     "cordis.patch.yml",
     "assets/architecture.svg",
     "assets/demo.gif",
     "scripts/prefetch.mjs",
     "scripts/list-voices.mjs",
+    "scripts/bench-asr.mjs",
+    "scripts/verify.mjs",
     "README.md",
     "README.en.md",
     "LICENSE"
@@ -36,10 +39,11 @@
   "scripts": {
     "build": "node build.mjs",
     "prepack": "node build.mjs",
-    "test": "node test/segmenter.test.mjs && node test/wakeword.test.mjs && node test/verify-client.mjs",
+    "test": "node test/segmenter.test.mjs && node test/aec.test.mjs && node test/download.test.mjs && node test/endpoint.test.mjs && node test/resample.test.mjs && node test/sense-worker.test.mjs && node test/detect-route.test.mjs && node test/verify-client.mjs",
     "verify:client": "node test/verify-client.mjs",
     "prefetch": "node scripts/prefetch.mjs",
     "list-voices": "node scripts/list-voices.mjs",
+    "bench-asr": "node scripts/bench-asr.mjs",
     "verify": "node scripts/verify.mjs",
     "typecheck": "node node_modules/typescript/bin/tsc -p tsconfig.json --noEmit && node node_modules/typescript/bin/tsc -p tsconfig.client.json --noEmit"
   },

package/scripts/bench-asr.mjs

@@ -0,0 +1,289 @@
+#!/usr/bin/env node
+/**
+ * P4-2 在线 ASR 换型离线对照评测（CER/段延迟/体积）。
+ *
+ * 用法：
+ *   node scripts/bench-asr.mjs --dir <测试集目录>
+ *
+ * 测试集目录约定：*.wav（16k 单声道 16bit PCM）+ 同名 *.txt（参考文本，UTF-8）。
+ * 模型自动懒下载至平台缓存目录（与插件同一约定：Linux/macOS
+ * ~/.cache/dsh-voice-mode/models/，Windows %LOCALAPPDATA%\dsh-voice-mode\models），
+ * .part 断点续传，huggingface.co ↗ hf-mirror.com 回退（--host 可指定镜像）。
+ *
+ * 输出：Markdown 表格（model | CER% | 平均段延迟 ms | 模型体积 MB），
+ * 「数据说话」支撑 P4 换型决策（plan.md §3-P4）。
+ */
+import { createWriteStream, readdirSync, readFileSync, statSync, mkdirSync, renameSync } from 'node:fs'
+import { dirname, join } from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { homedir } from 'node:os'
+import sherpa_onnx from 'sherpa-onnx'
+
+const here = dirname(fileURLToPath(import.meta.url))
+const root = join(here, '..')
+
+// ---------- 模型清单（repo/文件清单/设备配置） ----------
+const MODELS = [
+  {
+    id: 'zipformer-zh-int8',
+    repo: 'csukuangfj/sherpa-onnx-streaming-zipformer-zh-int8-2025-06-30',
+    files: ['encoder.int8.onnx', 'decoder.onnx', 'joiner.int8.onnx', 'tokens.txt'],
+    make: (t) => ({
+      modelConfig: {
+        transducer: { encoder: t('encoder.int8.onnx'), decoder: t('decoder.onnx'), joiner: t('joiner.int8.onnx') },
+        tokens: t('tokens.txt'), numThreads: 4, provider: 'cpu', debug: 0,
+      },
+      decodingMethod: 'greedy_search',
+    }),
+  },
+  {
+    id: 'zipformer-zh-xlarge-int8',
+    repo: 'csukuangfj/sherpa-onnx-streaming-zipformer-zh-xlarge-int8-2025-06-30',
+    files: ['encoder.int8.onnx', 'decoder.onnx', 'joiner.int8.onnx', 'tokens.txt'],
+    make: (t) => ({
+      modelConfig: {
+        transducer: { encoder: t('encoder.int8.onnx'), decoder: t('decoder.onnx'), joiner: t('joiner.int8.onnx') },
+        tokens: t('tokens.txt'), numThreads: 4, provider: 'cpu', debug: 0,
+      },
+      decodingMethod: 'greedy_search',
+    }),
+  },
+  {
+    id: 'zipformer-small-ctc-zh-int8',
+    repo: 'csukuangfj/sherpa-onnx-streaming-zipformer-small-ctc-zh-int8-2025-04-01',
+    files: ['model.int8.onnx', 'tokens.txt'],
+    make: (t) => ({
+      modelConfig: {
+        zipformer2Ctc: { model: t('model.int8.onnx') },
+        tokens: t('tokens.txt'), numThreads: 4, provider: 'cpu', debug: 0,
+      },
+    }),
+  },
+  {
+    id: 'paraformer-bilingual-zh-en',
+    repo: 'csukuangfj/sherpa-onnx-streaming-paraformer-bilingual-zh-en',
+    files: ['encoder.int8.onnx', 'decoder.int8.onnx', 'tokens.txt'],
+    make: (t) => ({
+      modelConfig: {
+        paraformer: { encoder: t('encoder.int8.onnx'), decoder: t('decoder.int8.onnx') },
+        tokens: t('tokens.txt'), numThreads: 4, provider: 'cpu', debug: 0,
+      },
+    }),
+  },
+]
+
+// ---------- 参数 ----------
+function parseArgs(argv) {
+  const args = { dir: null, host: 'https://huggingface.co' }
+  for (let i = 0; i < argv.length; i++) {
+    if (argv[i] === '--dir') args.dir = argv[i + 1]
+    else if (argv[i] === '--host') args.host = argv[i + 1]
+  }
+  return args
+}
+
+// ---------- 平台缓存目录 ----------
+function cacheDir() {
+  return process.platform === 'win32'
+    ? join(process.env.LOCALAPPDATA ?? join(homedir(), 'AppData', 'Local'), 'dsh-voice-mode', 'models')
+    : join(homedir(), '.cache', 'dsh-voice-mode', 'models')
+}
+
+// ---------- 懒下载（.part 续传 + host 回退，与插件 ensureFile 同构） ----------
+async function ensureFile(repoDir, file, hosts) {
+  const localPath = join(repoDir, file)
+  try {
+    if (statSync(localPath).isFile()) return true
+  } catch {
+    // 缺失
+  }
+  if (file === 'tokens.txt') console.log(`  下载 ${file}…`)
+  else console.log(`  下载 ${file}（可能较大）…`)
+  mkdirSync(repoDir, { recursive: true })
+  const partPath = localPath + '.part'
+  let partSize = 0
+  try {
+    partSize = statSync(partPath).size
+  } catch {
+    // 无 .part
+  }
+  for (const host of hosts) {
+    try {
+      const url = `${host}/${repoDir.split(/[\\/]/).pop()}/resolve/main/${file}`
+      const headers = { 'user-agent': 'dsh-voice-mode-bench' }
+      if (partSize > 0) headers.range = `bytes=${partSize}-`
+      const res = await fetch(url, { headers })
+      if (res.status === 416) {
+        renameSync(partPath, localPath)
+        return true
+      }
+      if (res.status !== 200 && res.status !== 206) continue
+      const sink = createWriteStream(partPath, partSize > 0 ? { flags: 'a' } : {})
+      const reader = res.body.getReader()
+      for (;;) {
+        const { done, value } = await reader.read()
+        if (done) break
+        if (!sink.write(value)) await new Promise((r) => sink.once('drain', r))
+      }
+      await new Promise((resolve, reject) => {
+        sink.end(() => resolve())
+        sink.on('error', reject)
+      })
+      renameSync(partPath, localPath)
+      return true
+    } catch {
+      partSize = 0 // 换 host 重来
+    }
+  }
+  return false
+}
+
+
+async function ensureModels(cache, repo, files, hosts) {
+  const repoDir = join(cache, repo)
+  for (const f of files) {
+    if (!(await ensureFile(repoDir, f, hosts))) {
+      console.error(`  模型下载失败: ${repo}/${f}`)
+      return false
+    }
+  }
+  return true
+}
+
+// ---------- WAV 读取（readWaveFromBinaryData → {samples, sampleRate}） ----------
+async function loadTestSet(dir) {
+  const entries = readdirSync(dir)
+  const wavs = entries.filter((n) => n.endsWith('.wav')).sort()
+  const cases = []
+  for (const w of wavs) {
+    const txt = w.replace(/\.wav$/, '.txt')
+    if (!entries.includes(txt)) {
+      console.warn(`跳过 ${w}：无同名 ${txt} 参考文本`)
+      continue
+    }
+    const buf = readFileSync(join(dir, w))
+    let wav
+    try {
+      wav = sherpa_onnx.readWaveFromBinaryData(new Uint8Array(buf.buffer, buf.byteOffset, buf.byteLength))
+    } catch (e) {
+      console.warn(`跳过 ${w}：无法解析 WAV（${String(e).slice(0, 80)}）`)
+      continue
+    }
+    if (wav.sampleRate !== 16000) {
+      console.warn(`跳过 ${w}：采样率 ${wav.sampleRate}Hz ≠ 16k（请先重采样）`)
+      continue
+    }
+    const ref = readFileSync(join(dir, txt), 'utf8').trim()
+    cases.push({ name: w, samples: wav.samples, ref })
+  }
+  return cases
+}
+
+// ---------- 编辑距离（字符级，中文按字） ----------
+function cer(hyp, ref) {
+  const a = [...hyp]
+  const b = [...ref]
+  const m = a.length
+  const n = b.length
+  const dp = new Uint32Array((m + 1) * (n + 1))
+  for (let i = 0; i <= m; i++) dp[i * (n + 1)] = i
+  for (let j = 0; j <= n; j++) dp[j] = j
+  for (let i = 1; i <= m; i++) {
+    for (let j = 1; j <= n; j++) {
+      const cost = a[i - 1] === b[j - 1] ? 0 : 1
+      dp[i * (n + 1) + j] = Math.min(
+        dp[(i - 1) * (n + 1) + j] + 1,
+        dp[i * (n + 1) + j - 1] + 1,
+        dp[(i - 1) * (n + 1) + j - 1] + cost,
+      )
+    }
+  }
+  const dist = dp[m * (n + 1) + n]
+  const denom = Math.max(n, 1)
+  return (dist / denom) * 100
+}
+
+
+
+// ---------- 单模型评测 ----------
+async function evalModel(model, cases, hosts, cache) {
+  const repoDir = join(cache, model.repo)
+  if (!(await ensureModels(cache, model.repo, model.files, hosts))) {
+    return { id: model.id, cer: null, ms: null, mb: null, error: '模型下载失败' }
+  }
+  const t = (f) => join(repoDir, f)
+  const rec = sherpa_onnx.createOnlineRecognizer(model.make(t))
+  let distSum = 0
+  let lenSum = 0
+  const timings = []
+  for (const c of cases) {
+    const stream = rec.createStream()
+    const t0 = performance.now()
+    stream.acceptWaveform(16000, c.samples)
+    while (rec.isReady(stream)) rec.decode(stream)
+    const text = rec.getResult(stream).text
+    timings.push(performance.now() - t0)
+    distSum += cer(text, c.ref) * Math.max(c.ref.length, 1)
+    lenSum += Math.max(c.ref.length, 1)
+    stream.free?.()
+  }
+  rec.free?.()
+  let bytes = 0
+  for (const f of model.files) {
+    try {
+      bytes += statSync(join(repoDir, f)).size
+    } catch {
+      // ignore
+    }
+  }
+  return {
+    id: model.id,
+    cer: lenSum > 0 ? distSum / lenSum : null,
+    ms: timings.length > 0 ? timings.reduce((a, b) => a + b, 0) / timings.length : null,
+    mb: bytes / (1024 * 1024),
+  }
+}
+
+async function main() {
+  const args = parseArgs(process.argv.slice(2))
+  if (!args.dir) {
+    console.error('用法: node scripts/bench-asr.mjs --dir <测试集目录> [--host 镜像]')
+    process.exit(1)
+  }
+  const cache = cacheDir()
+  const hosts = [...new Set([args.host, 'https://huggingface.co', 'https://hf-mirror.com'].filter(Boolean))]
+  const cases = await loadTestSet(args.dir)
+  if (cases.length === 0) {
+    console.error(`测试集为空：${args.dir}（需 16k 单声道 16bit PCM .wav + 同名 .txt）`)
+    process.exit(1)
+  }
+  console.log(`测试集 ${args.dir}：${cases.length} 段（合计约 ${Math.round(cases.reduce((a, c) => a + c.samples.length, 0) / 16000)}s 音频）`)
+  console.log('')
+  const rows = []
+  for (const m of MODELS) {
+    const r = await evalModel(m, cases, hosts, cache)
+    rows.push(r)
+    console.log(
+      r.error
+        ? `- ${m.id}: ${r.error}`
+        : `- ${m.id}: CER ${r.cer.toFixed(2)}% · 平均段延迟 ${r.ms.toFixed(0)}ms · 体积 ${r.mb.toFixed(0)}MB`,
+    )
+  }
+  console.log('')
+  console.log('| 模型 | CER% | 平均段延迟 ms | 体积 MB |')
+  console.log('| --- | --- | --- | --- |')
+  for (const r of rows) {
+    console.log(
+      r.error || r.cer === null
+        ? `| ${r.id} | — | — | — (${r.error ?? '无数据'}) |`
+        : `| ${r.id} | ${r.cer.toFixed(2)} | ${r.ms.toFixed(0)} | ${r.mb.toFixed(0)} |`,
+    )
+  }
+  console.log('')
+  console.log('说明：CER = 字符编辑距离/参考长度；段延迟 = 整段喂入到 getResult 的墙钟；体积 = 模型文件磁盘占用。')
+}
+
+main().catch((e) => {
+  console.error('bench failed:', e)
+  process.exit(1)
+})

package/scripts/prefetch.mjs

@@ -38,26 +38,30 @@ async function downloadFile(repoDir, file) {
     return true
   }
   const partPath = `${localPath}.part`
-  const partSt = await stat(partPath).catch(() => null)
   for (const host of [HOST_PRIMARY, HOST_FALLBACK]) {
     const url = `${host}/${MODEL_REPO}/resolve/main/${file}`
     const headers = { 'user-agent': 'dsh-voice-mode/prefetch' }
-    const resumeFrom = partSt?.size ?? 0
+    // 续传基准每 host 重读：前一 host 可能已追加过 .part，复用过期大小会拼坏文件。
+    const resumeFrom = (await stat(partPath).catch(() => null))?.size ?? 0
     if (resumeFrom > 0) headers.range = `bytes=${resumeFrom}-`
     try {
-      const res = await fetch(url, { headers })
+      // 与 src/asr-host.ts 一致：15 分钟超时（60s 会掐断 161MB encoder 大模型）+ 完整性核对。
+      const res = await fetch(url, { headers, signal: AbortSignal.timeout(900000) })
       if (res.status === 416) {
         await rename(partPath, localPath).catch(() => undefined)
         console.log(`  ✓ ${file}（续传完成）`)
         return true
       }
       if (res.status !== 200 && res.status !== 206) continue
-      const total = Number(res.headers.get('content-length') ?? 0) + resumeFrom
+      // 仅 206 才续传；带 .part 却回 200 全量（CDN 忽略 Range）必须从头重写，
+      // 否则在旧字节上追加全量 → 半旧半新损坏（与 asr-host.ts 同款修复）。
+      const resume = res.status === 206 ? resumeFrom : 0
+      const total = Number(res.headers.get('content-length') ?? 0) + resume
       const statusMax = file.length + 18
-      const sink = createWriteStream(partPath, resumeFrom > 0 ? { flags: 'a' } : {})
+      const sink = createWriteStream(partPath, resume > 0 ? { flags: 'a' } : {})
       const reader = res.body
       if (!reader) continue
-      let received = resumeFrom
+      let received = resume
       let lastPct = -1
       for (;;) {
         const { done, value } = await reader.read()
@@ -71,8 +75,17 @@ async function downloadFile(repoDir, file) {
         }
       }
       await new Promise((resolve, reject) => {
-        sink.end(resolve)
+        // 仅当字节数与声明一致（或无声明且收到 EOF）才算成功：截断即失败换 host。
+        sink.on('finish', () => {
+          if (total > 0 && received < total) {
+            sink.destroy(new Error('download truncated'))
+            reject(new Error('download truncated'))
+          } else {
+            resolve(true)
+          }
+        })
         sink.on('error', reject)
+        sink.end()
       })
       await rename(partPath, localPath)
       process.stdout.write(`\r  ${file.padEnd(statusMax)} 100%\n`)

package/scripts/verify.mjs

@@ -0,0 +1,34 @@
+#!/usr/bin/env node
+/**
+ * 离线纯逻辑验证聚合（官方验证金字塔第 3 层）。
+ * 逐个执行无网络单测与清单自检；任一失败即非零退出。
+ * 用法：node scripts/verify.mjs（或 npm run verify）
+ */
+import { spawnSync } from 'node:child_process'
+import { dirname, join } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const here = dirname(fileURLToPath(import.meta.url))
+const root = join(here, '..')
+const tests = [
+  ['segmenter 单测', ['node', 'test/segmenter.test.mjs']],
+  ['aec 单测', ['node', 'test/aec.test.mjs']],
+  ['下载完整性 单测', ['node', 'test/download.test.mjs']],
+  ['语义端点 单测', ['node', 'test/endpoint.test.mjs']],
+  ['重采样 单测', ['node', 'test/resample.test.mjs']],
+  ['SenseVoice worker 协议 单测', ['node', 'test/sense-worker.test.mjs']],
+  ['打断检测路由 单测', ['node', 'test/detect-route.test.mjs']],
+  ['清单/产物自检', ['node', 'test/verify-client.mjs']],
+]
+let failed = 0
+for (const [name, cmd] of tests) {
+  const r = spawnSync(cmd[0], cmd.slice(1), { cwd: root, encoding: 'utf8' })
+  const tail = r.stdout.trim().split('\n').slice(-2).join(' | ')
+  console.log(`— ${name}: ${r.status === 0 ? 'PASS' : 'FAIL'}${tail ? `（${tail}）` : ''}`)
+  if (r.status !== 0) {
+    console.error(r.stderr || r.stdout)
+    failed++
+  }
+}
+console.log(failed === 0 ? '\nverify: 全部通过' : `\nverify: ${failed} 项失败`)
+process.exitCode = failed === 0 ? 0 : 1