dsh-hold-to-talk-plus

Hold-to-talk voice input for the DeepSeek Harness (DSH) Web composer — now with third-party ASR, a user wordbank, draft-as-context prompting, a settings page and live preview.

Hold the mouse on the composer, speak, release: the transcript drops into your draft. The gesture, the floating overlay, the "slide up to cancel" interaction and the 60 s auto-finalize are unchanged from upstream; everything below is what this fork adds.

中文详细文档见 README.zh.md。

Features

Feature	Upstream dsh-hold-to-talk@0.1.4	This fork
Local offline ASR (SenseVoice via sherpa-onnx)	✅	✅
Third-party cloud ASR (OpenAI-compatible /v1/audio/transcriptions, or Alibaba BaiLian chat/completions + input_audio)	❌	✅
Auto-detect protocol from endpoint URL (no more 400 on mismatch)	❌	✅
User wordbank (glossary + correct=wrong1|wrong2 correction rules)	❌	✅
Composer draft sent as context prompt	❌	✅
Settings page (Settings → Voice input) with provider presets and guided API-key acquisition	❌	✅
Connection test without saving (401/403/404/429/timeout translated to plain language)	❌	✅
Live preview while speaking on cloud ASR (re-decode of the tail window every 2 s)	❌	✅
Automatic fallback to local ASR when the cloud call fails	❌	✅

Install (as a DSH profile plugin)

This repo is a drop-in replacement for dsh-hold-to-talk, so it keeps the same package name and bundle id.

# 1) clone somewhere stable
git clone https://github.com/lcx19970714/dsh-hold-to-talk-plus.git "H:\DSH-鼠标语音输入"
cd "H:\DSH-鼠标语音输入"

# 2) dependencies (sherpa-onnx native runtime)
pnpm install

# 3) point your DSH desktop profile at the local copy
#    edit  %USERPROFILE%\.dsh\profiles\desktop\package.json
#    "dsh-hold-to-talk": "link:H:/DSH-鼠标语音输入"
#    then:
$node = "$env:USERPROFILE\.dsh\dsh-runtimes\dsh-primary-runtime\dependencies\node\bin\node.exe"
$pnpm = "$env:USERPROFILE\.dsh\dsh-runtimes\dsh-primary-runtime\dependencies\pnpm\bin\pnpm.cjs"
& $node $pnpm install --dir "$env:USERPROFILE\.dsh\profiles\desktop"

# 4) add plugin config to  %USERPROFILE%\.dsh\profiles\desktop\cordis.patch.yml
#    - id: dsh-hold-to-talk
#      name: dsh-hold-to-talk
#      config: { ... }   # see README.zh.md

# 5) fully quit DSH (tray → Quit) and reopen

Then open Settings → Voice input, pick a provider, paste an API key, hit Test connection, save.

Providers

Provider	Endpoint	Model	Protocol
Alibaba Cloud Model Studio (BaiLian)	https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions	qwen3-asr-flash	chat-audio
SiliconFlow (free tier)	https://api.siliconflow.cn/v1/audio/transcriptions	TeleAI/TeleSpeechASR	transcriptions
Self-hosted (faster-whisper / whisper.cpp / FunASR gateway)	http://127.0.0.1:8000/v1/audio/transcriptions	whisper-large-v3	transcriptions
Local offline (no key)	—	SenseVoiceSmall int8	local

Note: on BaiLian's compatible endpoint the ASR task ignores instruction-style prompts; the glossary is accepted but has limited effect. Reliable term correction comes from the wordbank's correct=wrong rules, which apply to both local and cloud results.

Layout

lib/
├─ index.js           # host: routes, engine management, prompt assembly, cloud branch, /settings
├─ client.js          # browser: hold gesture, overlay, settings page
├─ asr-worker.mjs     # local sherpa-onnx decoder (upstream)
├─ model-cache.js     # local model download cache (upstream)
├─ cloud-asr.mjs      # WAV encoding + two third-party protocols + system prompt
├─ wordbank.mjs       # glossary / correction rules parsing + local post-processing
└─ settings-store.mjs # settings persistence (~/.dsh/hold-to-talk/settings.json)

Credits & license

Fork of wangzhanchao883/dsh-hold-to-talk (MIT). Original copyright © 2026 wangzhanchao883; modifications © 2026 lcx19970714. Licensed under the MIT License — see LICENSE.