dsh-hold-to-talk

Hold-to-talk voice input for the DeepSeek Harness Web composer — the WeChat-desktop gesture: hold the mouse on the input box, speak, see the running transcript float above it, release to drop the text into the draft. Recognition runs locally (SenseVoice via sherpa-onnx): no API key, offline, audio never leaves the machine.

DSH Web 输入框的"长按说话"语音输入 —— 微信电脑版同款操作：在输入框上按住鼠标说话，浮层里边说边出字，松手把文字写进输入框，上滑取消。识别全部在本机跑（SenseVoice + sherpa-onnx），免密钥、离线、音频不出本机。

┌─ composer ──────────────────────────────────┐
│  [hold the mouse, don't move, 400ms]         │
│         ↓                                    │
│   ┌──────────────────────────────────┐       │
│   │ ●  ▁▃▅▂▇▃▁  3s                   │  ← overlay, fixed above the composer
│   │ release to send · slide up to cancel │    │
│   │ 开饭时间早上9点至下午5点            │  ← live transcript (preview only)
│   └──────────────────────────────────┘       │
│  [release] → final text appended to the draft │
└──────────────────────────────────────────────┘

English · 中文说明

Demo / 演示

按住鼠标说话 → 浮层里边说边出字 → 松手把文字写进输入框（35 秒实录，完整视频）

1. Hold on the input box / 长按进入录音	2. Release, text lands in the draft / 松手写入草稿	3. Plugin card / 插件卡片
The overlay appears with "release to send · slide up to cancel" / 浮层出现，提示"松开发送 · 上滑取消"	The transcript is appended to the composer; the live preview stayed in the overlay / 识别文字写入输入框（边说边出字的预览只留在浮层里）	Listed in the plugin list after install / 安装后在插件列表可见

English

What it does

• A long press on the composer starts dictation; releasing inserts the transcript into the draft; sliding up cancels; Esc cancels; a 60-second hold auto-finishes.
• While you speak, an overlay above the composer shows a live transcript, a level meter and the elapsed time.
• The live transcript is a preview only — it is never written into the draft. Only the release-time result (decoded from the full audio) is inserted, so recognition jitter cannot corrupt what you are editing.
• Recognition is local: SenseVoice (int8) through sherpa-onnx, decoded in a worker thread. No API key, no cloud call, no telemetry.
• Normal editing is untouched: a quick click still places the caret, dragging still selects text, and holding without reaching the threshold does nothing.

Interaction

Requirements

• DeepSeek Harness running the web profile (dsh web).
• Node.js >=22.19 or >=24 on the machine running DSH.
• A Chromium-based browser (Chrome / Edge) with microphone access. The DSH web UI is served over http://127.0.0.1, which is a secure context, so getUserMedia is allowed.
• ~230 MB of disk for the model, downloaded once on first use.

Install

dsh plugin --profile web add dsh-hold-to-talk
# or from a checkout / GitHub:
dsh plugin --profile web add /path/to/dsh-hold-to-talk
# then restart dsh web and hard-refresh the page (Ctrl+Shift+R)

First run

On page load the plugin starts downloading the model in the background (model.int8.onnx, 228 MB, from hf-mirror.com, resumable). While it downloads, holding the composer shows "语音模型准备中 x%". To fetch it ahead of time:

npm run model:fetch

The model is cached under ~/.dsh/hold-to-talk/models/ (change with modelDir).

Configuration

Namespace dsh-hold-to-talk (written to settings.yaml by the Web settings panel), or override in the profile's cordis.patch.yml:

- id: dsh-hold-to-talk
  name: dsh-hold-to-talk
  config:
    holdThresholdMs: 400     # long-press threshold
    cancelSlidePx: 60        # slide-up-to-cancel distance
    interimIntervalMs: 1500  # live-preview cadence
    maxWindowSec: 6          # preview decode window; the final always uses all audio
    minHoldMs: 250           # shorter holds are discarded (mis-taps)
    maxHoldMs: 60000         # auto-finish limit
    autoSend: false          # insert only; never auto-send by default
    language: auto           # auto / zh / en / ja / ko / yue
    useItn: true             # inverse text normalization ("二零二六" -> "2026")
    numThreads: 2            # native engine threads
    mirror: https://hf-mirror.com
    modelDir: ""             # empty = ~/.dsh/hold-to-talk/models

Priority: settings panel > cordis.patch.yml > plugin defaults.

How it works

browser half (lib/client.js)                    host half (lib/index.js)
┌───────────────────────────────┐   HTTP      ┌────────────────────────────────┐
│ gesture  capture-phase events  │             │ exact routes via webServer      │
│  mousedown→400ms→move/select   │             │  POST /asr  ?mode=interim|final │
│ capture  AudioWorklet→16k PCM  │────────────▶│             |drop              │
│  ring buffer / linear resample │             │  GET  /health[?prepare=1]       │
│ overlay  conversation.input.   │◀────────────│  GET  /config                   │
│          overlay slot          │   JSON      │ per-hold buffer: interim decodes│
│ insert   inputActions.setDraft  │            │ the tail window, final uses all │
└───────────────────────────────┘             │ worker_threads ↓                │
                                               │ sherpa-onnx SenseVoice (int8)   │
                                               └────────────────────────────────┘

Engineering decisions worth knowing:

• Decoding runs in a worker thread. sherpa-onnx decoding is synchronous and blocks for 0.6–3.3 s per 6–22 s of audio; on the main thread it would stall the host event loop and the page's streaming output with it.
• Native first, WASM fallback. sherpa-onnx-node (native, multi-threaded) is ~2.5× faster than sherpa-onnx (WASM, single-threaded) — see the table below. WASM lives in optionalDependencies; the worker falls back automatically.
• Absolute model paths only. The WASM build uses NODERAWFS; a relative path lands in emscripten's virtual CWD and is not found.
• The AudioContext is created inside the mousedown gesture. Creating it 400 ms later gets it suspended by autoplay policy and no audio flows. The capture node feeds a zero-gain sink before destination so it is pulled without being audible (otherwise: feedback howl).
• Resampling always restarts from sample 0, so successive increments tile the stream exactly; resampling only the tail would drift.
• Previews wait for 2 seconds of audio. A non-streaming model fed too little hallucinates (1.2 s produced "哈。", 0.5 s produced "Yeah.").
• Late previews are dropped by epoch, so a preview arriving after release can never overwrite the final transcript.
• Gesture ownership is decided by DOM containment, never by visibility. The overlay is hidden while idle; dispatching by offsetParent/getClientRects() made long-press die permanently after the first successful dictation (v0.1.0).
• The overlay anchor is always rendered (zero-size, out of flow) so the ownership anchor never disappears, and a module token makes the previous module's document listeners inert after a hot reload.

Measured performance

Decode time on the author's machine (Windows, Node 22.22):

Model load into the worker happens once: 2158 ms (WASM) / 3565 ms (native).

Limitations

• SenseVoice is not a streaming model. "Live" is a re-decode of the tail window every 1.5 s, so the preview trails speech by roughly 1.5–2 s; the final result (on release) always uses the complete audio.
• Long dictation takes a moment to finish: ~3 s for 20 s of speech, ~9 s at the 60 s cap.
• Desktop mouse only. Touch long-press is deliberately not implemented — it collides with text selection and the context menu on mobile.
• The model download is 228 MB on first use.

Development

npm run check            # syntax check all runtime modules
npm test                 # check + three offline suites
npm run model:fetch      # 228MB model (needed by the full suites)
npm run bench            # decode-time benchmark

None of the suites need a browser or a DSH restart:

Without the model, the model-dependent cases skip themselves, so npm test is green on a clean clone — which is exactly what CI runs.

Security

• Audio never leaves the machine. It travels only over loopback HTTP to the plugin's own route and is decoded by a local process. No third-party service is contacted during recognition.
• No credentials. The plugin reads and stores no API key or token.
• Routes refuse cross-site requests (sec-fetch-site), so a random web page cannot post audio at the local server.
• The microphone is opened only while you hold; the stream's tracks are stopped on release, and no other page or process gets the audio.
• The only network request the plugin makes by itself is the one-time model download from the configured mirror (hf-mirror.com by default). No telemetry.

License

MIT — see LICENSE.

中文说明

给 DeepSeek Harness Web 界面加微信电脑版同款语音输入：在输入框上按住鼠标说话，浮层里边说边出字，松手把文字写进输入框，按住上滑则丢弃。识别在你自己的机器上跑（SenseVoice + sherpa-onnx），不需要任何 API key，音频不出本机。

功能

• 输入框上长按鼠标即进入语音输入，松手把识别结果追加进输入框；上滑取消、Esc 取消、按满 60 秒自动定稿。
• 说话期间输入框上方浮层显示实时字幕、音量波形与计时。
• 实时字幕只作预览，绝不回写输入框；只有松手时用完整音频解码出的结果才写入，识别抖动不会污染你正在编辑的草稿。
• 识别在本机完成（SenseVoice int8 + sherpa-onnx，解码放在 worker 线程）：免密钥、离线、无遥测。
• 不影响正常编辑：快速单击照旧放光标、拖动照旧选中文字，没到长按阈值就松手什么都不会发生。

交互细节

两个刻意的设计取舍：

1. 实时预览只画在浮层里，绝不回写输入框 —— 识别抖动/错字不会污染你正在编辑的草稿；只有松手后的定稿才写入。
2. 松手才出字（与微信一致，微信也是"松开后转为文字"）。"边说边出字"是加在浮层上的增强。

环境要求

• DeepSeek Harness，运行 web profile（dsh web）。
• 运行 DSH 的机器上 Node.js >=22.19 或 >=24。
• Chromium 系浏览器（Chrome / Edge）并允许麦克风。DSH Web 走 http://127.0.0.1，属安全上下文，getUserMedia 可用。
• 首次使用需要约 230MB 磁盘放模型。

安装

dsh plugin --profile web add dsh-hold-to-talk
# 或从本地目录 / GitHub:
dsh plugin --profile web add /path/to/dsh-hold-to-talk
# 装完重启 dsh web,然后 Ctrl+Shift+R 硬刷新页面

⚠️ Windows 上跑 dsh CLI 要用 DSH 自己那个 Node。CLI 入口靠 import.meta.main 自执行（Node 22.18+/24 才有），在旧版 Node 上会静默退出、什么都不做——看起来安装成功了，实际 profile 里没有任何变化。用运行中的 DSH 同款解释器：

& 'D:\Program Files\QClaw\<版本>\resources\node\node.exe' `
  'C:\Users\Administrator\AppData\Roaming\QClaw\npm-global\node_modules\@deepseek-ai\dsh\lib\bin.js' `
  plugin --profile web add D:\workout\deepseekharness\dsh-hold-to-talk

首次使用

页面加载后插件会在后台自动下载模型（model.int8.onnx 228MB + tokens.txt，走 hf-mirror.com，支持断点续传）。下载期间长按会看到"语音模型准备中 x%"，只会发生一次。想提前下好：

npm run model:fetch

模型缓存在 ~/.dsh/hold-to-talk/models/（可用 modelDir 改）。

配置

设置命名空间 dsh-hold-to-talk（Web 设置面板写入 settings.yaml），或在 profile 的 cordis.patch.yml 里按 id 覆盖：

- id: dsh-hold-to-talk
  name: dsh-hold-to-talk
  config:
    holdThresholdMs: 400     # 长按判定阈值
    cancelSlidePx: 60        # 上滑取消的位移
    interimIntervalMs: 1500  # 实时预览间隔
    maxWindowSec: 6          # 预览解码的滑动窗口上限;定稿始终用完整音频
    minHoldMs: 250           # 短于此时长直接丢弃(防误触)
    maxHoldMs: 60000         # 单次长按上限,到点自动定稿
    autoSend: false          # 定稿后是否自动发送(默认只写进输入框)
    language: auto           # auto / zh / en / ja / ko / yue
    useItn: true             # 逆文本规整("二零二六年"→"2026年")
    numThreads: 2            # 原生引擎线程数
    mirror: https://hf-mirror.com
    modelDir: ""             # 留空 = ~/.dsh/hold-to-talk/models

优先级：设置面板 > cordis.patch.yml > 插件默认值。

使用

1. 打开任意会话，把鼠标停在输入框上按住不动约 0.4 秒（首次会弹麦克风授权，允许）。
2. 浮层出现后开始说话，字幕会一段一段跟上来。
3. 松手 → 定稿文字追加进输入框，由你自己按发送。
4. 说错了就按住上滑再松手，或按 Esc 直接丢弃。

目录结构

dsh-hold-to-talk/
├── package.json              # dsh.bundle / dsh.client 声明、依赖、脚本
├── cordis.patch.yml          # 宿主行注册 + 全部默认配置
├── lib/
│   ├── index.js              # 宿主:三条 HTTP 路由 + 引擎管理 + 模型下载 + 会话缓冲 + 设置命名空间
│   ├── asr-worker.mjs        # worker 线程内解码(原生优先、WASM 兜底)
│   ├── model-cache.js        # 模型下载缓存(镜像、断点续传、进度)
│   └── client.js             # 浏览器:长按手势 + AudioWorklet 采集 + 浮层 + 上屏
├── tools/                    # 离线自测与开发脚本(不随 npm 包发布)
│   ├── client-smoke.mjs      # 客户端注册契约
│   ├── host-test.mjs         # 真实 HTTP 路由端到端
│   ├── client-test.mjs       # 在 Node 里驱动真实交互逻辑
│   ├── decode-test.mjs       # 模型+worker+识别冒烟
│   ├── bench.mjs             # 解码耗时基准
│   ├── fetch-model.mjs       # 拉模型
│   ├── market-note.mjs       # 写 DSH 市场卡片备注的小工具
│   ├── market-note.txt       # 上面的文案
│   └── zh.wav                # 测试音频(k2-fsa sherpa-onnx 示例音频,Apache-2.0)
├── .github/workflows/test.yml
├── LICENSE
├── README.md
├── CHANGELOG.md
└── CONTRIBUTING.md

架构速览

浏览器半身 (lib/client.js)                     宿主半身 (lib/index.js)
┌───────────────────────────────┐   HTTP      ┌────────────────────────────────┐
│ 手势层  document 捕获阶段        │             │ webServer.register 精确路由       │
│  mousedown→400ms→移动/选区取消   │             │  POST /asr   ?mode=interim|final │
│ 采集层  AudioWorklet→16k f32 PCM │────────────▶│              |drop             │
│  环形缓冲/线性重采样             │             │  GET  /health[?prepare=1]        │
│ 浮层    conversation.input.       │◀────────────│  GET  /config                    │
│         overlay 槽               │   JSON      │ 会话缓冲:interim 解码尾部窗口,      │
│ 上屏    inputActions.setDraft()   │             │          final 用完整音频        │
└───────────────────────────────┘             │ worker_threads ↓                │
                                                │ sherpa-onnx SenseVoice(int8)     │
                                                └────────────────────────────────┘

几个关键工程决定：

• 解码放 worker 线程：sherpa-onnx 的解码是同步阻塞的，一次 6~22 秒音频要占住线程 0.6~3.3 秒。放主线程会拖住宿主事件循环、让页面流式输出跟着卡。
• 原生优先、WASM 兜底：sherpa-onnx-node（原生、可多线程）比 sherpa-onnx（WASM 单线程）快约 2.5 倍，实测见下表。WASM 放在 optionalDependencies 里，原生加载失败时 worker 自动降级。
• 模型路径必须绝对：sherpa-onnx 的 WASM 构建启用了 NODERAWFS，相对路径会落到 emscripten 的虚拟 CWD 里读不到（实测）。
• AudioContext 在 mousedown 的用户手势里预先创建：等到 400ms 后再建会被浏览器自动播放策略挂成 suspended，音频流不起来。采集节点串一个 0 增益 sink 再进 destination，既被拉动又不外放（否则啸叫）。
• 重采样每次从 0 号样本重算：只重采样尾部会产生错位，增量拼接就不连续了。
• interim 至少攒够 2 秒才解码：非流式模型喂太短会吐幻觉词（实测 1.2 秒得到"哈。"，0.5 秒得到 "Yeah."）。
• 迟到的 interim 用 epoch 丢弃：松手后到达的预览结果永远不能覆盖定稿。
• 长按归属用 DOM 包含关系判定，绝不看可见性：浮层空闲时是隐藏的，若用 offsetParent/getClientRects() 判"我这个实例可见吗"，会在第一次识别成功后长按彻底失效（v0.1.0 实测踩到这个坑）。改成"从输入框往上找同时包含浮层节点的祖先"，与 CSS 无关；单实例时兜底，多实例且都对不上宁可不响应（避免把文字写进别的会话）。
• 浮层锚点常驻：外层标记节点永远渲染（哪怕卡片隐藏），它是归属判定的 DOM 锚点，零尺寸、脱离文档流。
• HMR token：整包热重载后，上一份模块注册的文档级监听器自动失效，避免同一次 mousedown 被新旧两份逻辑各接一次。

实测性能

本机（Windows + Node 22.22）解码耗时：

模型加载（首次进入 worker）：WASM 2158ms / 原生 3565ms，只发生一次。

已知限制

• 不是流式模型：SenseVoice 是非流式识别，"实时"是每 1.5 秒重解码尾部窗口做出来的预览，因此预览比说话滞后约 1.5~2 秒；定稿（松手）永远用完整音频，结果最准。
• 长语音定稿要等：说 20 秒约等 3 秒。60 秒上限约 9 秒。
• 仅桌面鼠标：没有实现 touch 长按（移动端长按会与文本选择/上下文菜单冲突），故意留空。
• 麦克风需要安全上下文：http://127.0.0.1 属安全上下文，可用；若通过局域网 IP 访问需 https。
• 首次下载 228MB 模型；~/.dsh 所在盘需留出空间。

开发与自测

npm run check            # 四个运行模块的语法检查
npm test                 # 上面 + 三套离线测试
npm run model:fetch      # 拉模型(228MB,全量测试需要)
npm run bench            # 解码耗时基准

三套自测都不需要重启 DSH、不需要浏览器：

没有模型时依赖识别的用例会自动跳过，所以在干净克隆上 npm test 也是绿的——CI 跑的就是这个。

安全说明

• 音频不出本机：只在本机 loopback HTTP 上传输给自己插件的路由，由本地进程解码，识别过程不接触任何第三方服务。
• 不涉及任何凭据：插件不读取、不存储 API key 或 token。
• 路由拒绝跨站请求：按 sec-fetch-site 拦截，任意网页无法向本机服务投递音频。
• 麦克风只在按住期间打开：松手即停止音轨，其他页面/进程拿不到音频。
• 插件自身唯一的联网行为是首次按配置镜像（默认 hf-mirror.com）下载模型；无遥测。

卸载

dsh plugin --profile web remove dsh-hold-to-talk
# 重启 dsh web;模型缓存在 ~/.dsh/hold-to-talk/,可自行删除

License

MIT — 见 LICENSE。