214 lines
9.8 KiB
TypeScript
214 lines
9.8 KiB
TypeScript
/**
|
||
* 端侧 ASR 资产获取：pnpm fetch:asr
|
||
*
|
||
* 下载 sherpa-onnx WASM 运行时 + SenseVoiceSmall int8 模型，产物落到 public/asr/
|
||
* （该目录在 .gitignore 里，模型不进 git；换机器/删了目录就重跑本脚本）。
|
||
*
|
||
* 资产来源（2026-10-04 核实，大小为当天下载实测字节数）：
|
||
* - WASM 运行时：HuggingFace 官方 space k2-fsa/web-assembly-vad-asr-sherpa-onnx-zh-zipformer-ctc
|
||
* https://huggingface.co/spaces/k2-fsa/web-assembly-vad-asr-sherpa-onnx-zh-zipformer-ctc/resolve/main/<文件名>
|
||
* （glue 95,308 B / wasm 11,722,172 B / api 47,391 B）
|
||
* 注意：普通 wasm/asr 运行时（GitHub release 的 zipformer/paraformer demo）只导出
|
||
* Online（流式）API，Offline 系列不可用；vad-asr 构建才导出 OfflineRecognizer
|
||
* （wasm/vad-asr/CMakeLists.txt 的 exported_functions），而 SenseVoice 是离线模型，
|
||
* 必须用这一份。文件逐个下载，无需解压 tarball。
|
||
* - 模型：GitHub release asr-models
|
||
* https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17.tar.bz2
|
||
* （163,002,883 B；内含 model.int8.onnx 239,233,841 B + tokens.txt 315,894 B）
|
||
*
|
||
* 关键改造：官方 glue（sherpa-onnx-wasm-main-vad-asr.js）内嵌的 Emscripten 打包清单指向
|
||
* zipformer-ctc demo 的 silero_vad + zipformer 模型（367MB .data）。我们把清单改写为只含
|
||
* /model.int8.onnx + /tokens.txt 两个文件，并据此生成自己的 .data（两文件直接拼接，
|
||
* Emscripten preload 格式 = 原始字节流 + JS 侧清单）。这样运行时一次请求拿到模型包，
|
||
* 模型出现在 MEMFS 根路径，OfflineRecognizer 按绝对路径读取。
|
||
*
|
||
* 产物（public/asr/）：
|
||
* sherpa-onnx-wasm-main-vad-asr.js glue（清单已改写）
|
||
* sherpa-onnx-wasm-main-vad-asr.wasm 运行时 wasm（11,722,172 B，含 VAD + Offline API）
|
||
* sherpa-onnx-wasm-main-vad-asr.data 模型预加载包（≈239,549,735 B）
|
||
* sherpa-onnx-asr.js 官方 JS API 封装（OfflineRecognizer）
|
||
* sense-voice-LICENSE.txt 模型许可（FunASR 模型许可，允许商用需署名）
|
||
* manifest.json 资产清单 + 字节数（引擎下载进度与完整性自检用）
|
||
*/
|
||
import { spawnSync } from 'node:child_process';
|
||
import { createWriteStream } from 'node:fs';
|
||
import fs from 'node:fs/promises';
|
||
import os from 'node:os';
|
||
import path from 'node:path';
|
||
import { fileURLToPath } from 'node:url';
|
||
|
||
const ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');
|
||
const OUT_DIR = path.join(ROOT, 'public', 'asr');
|
||
const FORCE = process.argv.includes('--force');
|
||
|
||
const HF_SPACE =
|
||
'https://huggingface.co/spaces/k2-fsa/web-assembly-vad-asr-sherpa-onnx-zh-zipformer-ctc/resolve/main';
|
||
const GLUE_NAME = 'sherpa-onnx-wasm-main-vad-asr.js';
|
||
const WASM_NAME = 'sherpa-onnx-wasm-main-vad-asr.wasm';
|
||
const API_NAME = 'sherpa-onnx-asr.js';
|
||
const DATA_NAME = 'sherpa-onnx-wasm-main-vad-asr.data';
|
||
const RUNTIME_FILES: Array<{ name: string; bytes: number }> = [
|
||
{ name: GLUE_NAME, bytes: 95_308 },
|
||
{ name: WASM_NAME, bytes: 11_722_172 },
|
||
{ name: API_NAME, bytes: 47_391 },
|
||
];
|
||
|
||
const MODEL_URL =
|
||
'https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17.tar.bz2';
|
||
const MODEL_BYTES = 163_002_883;
|
||
// 模型 tarball 内
|
||
const MODEL_ONNX_NAME = 'model.int8.onnx';
|
||
const TOKENS_NAME = 'tokens.txt';
|
||
const LICENSE_NAME = 'sense-voice-LICENSE.txt';
|
||
|
||
async function download(url: string, dest: string, expectedBytes: number): Promise<void> {
|
||
for (let attempt = 1; attempt <= 3; attempt++) {
|
||
try {
|
||
const res = await fetch(url);
|
||
if (!res.ok || !res.body) throw new Error(`HTTP ${res.status}`);
|
||
const reader = res.body.getReader();
|
||
const out = createWriteStream(dest);
|
||
let received = 0;
|
||
let logged = 0;
|
||
for (;;) {
|
||
const { done, value } = await reader.read();
|
||
if (done) break;
|
||
out.write(value);
|
||
received += value.byteLength;
|
||
if (received - logged >= 20 * 1024 * 1024) {
|
||
logged = received;
|
||
console.log(` … ${(received / 1e6).toFixed(0)} MB`);
|
||
}
|
||
}
|
||
await new Promise((resolve, reject) => out.end(resolve as () => void, reject));
|
||
if (received !== expectedBytes) {
|
||
throw new Error(`大小不符：期望 ${expectedBytes}，实际 ${received}（上游资产可能已更新，请核对后改脚本常量）`);
|
||
}
|
||
return;
|
||
} catch (err) {
|
||
console.log(` 下载失败（第 ${attempt} 次）：${err instanceof Error ? err.message : err}`);
|
||
if (attempt === 3) throw err;
|
||
}
|
||
}
|
||
}
|
||
|
||
async function extract(tarball: string, dir: string): Promise<void> {
|
||
const r = spawnSync('tar', ['-xf', tarball, '-C', dir], { stdio: 'inherit' });
|
||
if (r.status !== 0) throw new Error(`解压失败：${tarball}`);
|
||
}
|
||
|
||
/** 只含一个顶层目录时返回该目录路径，否则返回 dir 本身 */
|
||
async function onlySubdir(dir: string): Promise<string> {
|
||
const entries = await fs.readdir(dir, { withFileTypes: true });
|
||
const dirs = entries.filter((e) => e.isDirectory());
|
||
return dirs.length === 1 && entries.length === 1 ? path.join(dir, dirs[0].name) : dir;
|
||
}
|
||
|
||
/** 把 glue 里 loadPackage({...}) 的实参整体替换为新的打包清单（括号配平扫描，兼容 key 带不带引号两种生成风格） */
|
||
export function patchGlueMetadata(glue: string, files: Array<{ filename: string; start: number; end: number }>, total: number): string {
|
||
const anchor = 'loadPackage({';
|
||
const start = glue.indexOf(anchor);
|
||
if (start === -1) throw new Error('glue 中找不到 loadPackage 调用');
|
||
let depth = 0;
|
||
for (let i = start + 'loadPackage('.length; i < glue.length; i++) {
|
||
const ch = glue[i];
|
||
if (ch === '{') depth++;
|
||
else if (ch === '}') {
|
||
depth--;
|
||
if (depth === 0) {
|
||
const meta = `{files:[${files.map((f) => `{filename:"${f.filename}",start:${f.start},end:${f.end}}`).join(',')}],remote_package_size:${total}}`;
|
||
return glue.slice(0, start) + 'loadPackage(' + meta + glue.slice(i + 1);
|
||
}
|
||
}
|
||
}
|
||
throw new Error('glue 中 loadPackage 清单括号不配平');
|
||
}
|
||
|
||
async function main() {
|
||
await fs.mkdir(OUT_DIR, { recursive: true });
|
||
|
||
const manifestPath = path.join(OUT_DIR, 'manifest.json');
|
||
if (!FORCE && (await fs.stat(manifestPath).catch(() => null)) !== null) {
|
||
console.log('public/asr/ 已存在（--force 可强制重新下载）');
|
||
return;
|
||
}
|
||
|
||
const tmp = await fs.mkdtemp(path.join(os.tmpdir(), 'jev-asr-'));
|
||
console.log(`临时目录：${tmp}`);
|
||
|
||
console.log('\n1/2 下载 WASM 运行时（vad-asr 构建，含 Offline API）');
|
||
for (const f of RUNTIME_FILES) {
|
||
console.log(` ${HF_SPACE}/${f.name}`);
|
||
await download(`${HF_SPACE}/${f.name}`, path.join(tmp, f.name), f.bytes);
|
||
}
|
||
|
||
console.log(`\n2/2 下载 SenseVoiceSmall int8 模型（${(MODEL_BYTES / 1e6).toFixed(0)} MB）\n ${MODEL_URL}`);
|
||
const modelTar = path.join(tmp, 'model.tar.bz2');
|
||
await download(MODEL_URL, modelTar, MODEL_BYTES);
|
||
const modelDir = path.join(tmp, 'model');
|
||
await fs.mkdir(modelDir);
|
||
await extract(modelTar, modelDir);
|
||
const model = await onlySubdir(modelDir);
|
||
|
||
const glue = await fs.readFile(path.join(tmp, GLUE_NAME), 'utf8');
|
||
const wasm = await fs.readFile(path.join(tmp, WASM_NAME));
|
||
const api = await fs.readFile(path.join(tmp, API_NAME));
|
||
const modelOnnx = await fs.readFile(path.join(model, MODEL_ONNX_NAME));
|
||
const tokens = await fs.readFile(path.join(model, TOKENS_NAME));
|
||
const licenseText = await fs.readFile(path.join(model, 'LICENSE')).catch(() => null);
|
||
|
||
// 生成 .data：model.int8.onnx 拼 tokens.txt；清单路径用绝对路径（MEMFS 根）
|
||
const data = Buffer.concat([modelOnnx, tokens]);
|
||
const patched = patchGlueMetadata(
|
||
glue,
|
||
[
|
||
{ filename: `/${MODEL_ONNX_NAME}`, start: 0, end: modelOnnx.byteLength },
|
||
{ filename: `/${TOKENS_NAME}`, start: modelOnnx.byteLength, end: data.byteLength },
|
||
],
|
||
data.byteLength,
|
||
);
|
||
if (patched === glue) throw new Error('清单改写失败');
|
||
|
||
const assets: Array<{ name: string; bytes: number }> = [
|
||
{ name: GLUE_NAME, bytes: Buffer.byteLength(patched) },
|
||
{ name: WASM_NAME, bytes: wasm.byteLength },
|
||
{ name: DATA_NAME, bytes: data.byteLength },
|
||
{ name: API_NAME, bytes: api.byteLength },
|
||
];
|
||
|
||
await fs.writeFile(path.join(OUT_DIR, GLUE_NAME), patched);
|
||
await fs.writeFile(path.join(OUT_DIR, WASM_NAME), wasm);
|
||
await fs.writeFile(path.join(OUT_DIR, DATA_NAME), data);
|
||
await fs.writeFile(path.join(OUT_DIR, API_NAME), api);
|
||
if (licenseText) await fs.writeFile(path.join(OUT_DIR, LICENSE_NAME), licenseText);
|
||
await fs.writeFile(
|
||
path.join(OUT_DIR, 'manifest.json'),
|
||
JSON.stringify(
|
||
{
|
||
generatedAt: new Date().toISOString(),
|
||
runtime: { source: `${HF_SPACE} （k2-fsa 官方 space，vad-asr 构建）`, note: 'wasm/asr 构建不含 Offline API，SenseVoice 必须用 vad-asr 运行时' },
|
||
model: { url: MODEL_URL, name: 'sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17' },
|
||
assets,
|
||
},
|
||
null,
|
||
2,
|
||
),
|
||
);
|
||
|
||
// 自检：清单字节数与落盘文件一致
|
||
for (const a of assets) {
|
||
const actual = (await fs.stat(path.join(OUT_DIR, a.name))).size;
|
||
if (actual !== a.bytes) throw new Error(`自检失败：${a.name} 大小 ${actual} != 清单 ${a.bytes}`);
|
||
}
|
||
|
||
await fs.rm(tmp, { recursive: true, force: true });
|
||
console.log(`\n完成。产物在 public/asr/（共 ${(assets.reduce((n, a) => n + a.bytes, 0) / 1e6).toFixed(1)} MB）：`);
|
||
for (const a of assets) console.log(` ${(a.bytes / 1e6).toFixed(2).padStart(8)} MB ${a.name}`);
|
||
console.log('\n首次在浏览器/桌面端使用时懒加载到 CacheStorage；见 src/asr/ondevice.ts 头注释。');
|
||
}
|
||
|
||
main().catch((err) => {
|
||
console.error(err);
|
||
process.exit(1);
|
||
});
|