FuzzySoul/dsh-chatvoice

ChatVoice 🎤🔊 — dsh-chatvoice

English | 中文

Free, zero-config, no-API-key voice for DeepSeek Harness (dsh): speak your prompts and have AI replies read aloud. Everything runs on the browser's native Web Speech API — no backend, no key, nothing to register.

_{🎤 Voice input: each confirmed sentence lands in the box in real time}

_{🔊 Read aloud: click the speaker on any reply, stop anytime}

_{✏️ Edit while listening: type fixes or delete while the mic runs — append-only, never rewrites}

ChatVoice = Chat + Voice: one plugin for both your mouth and your ears — dictate prompts while your hands stay on the keyboard, and let AI read long replies to you (listening-based learning, accessibility, or just lying back).

Features

Why Edge is recommended

Install

dsh plugin --profile web add dsh-chatvoice
# or manually: pnpm add dsh-chatvoice (dsh.profile.bundles reconciles automatically)

Restart dsh web (dsh web) and open http://127.0.0.1:3080.

⚠️ You must access dsh web via 127.0.0.1: speech recognition requires a secure context (HTTPS or localhost). Over a LAN IP the browser blocks the microphone — input is disabled with a hint, read-aloud still works.

Usage

1. Voice input: click 🎤 in the composer toolbar → allow the microphone permission → keep talking (each confirmed sentence lands in the box in real time, interim results show in the bubble) → click 🎤 again to stop → press Enter to send. While listening you can type fixes or clear the box entirely — speech only appends to the end and never rewrites, so nothing you deleted comes back
2. Read aloud: click 🔊 next to an assistant reply → it reads aloud (button turns into a red ⏹) → click again to stop
3. Auto-read: Settings → ChatVoice → enable "Auto-read new replies" → save; new replies are read automatically

Settings

How it works

• host (dsh/index.js): Config schema + GET/POST /dsh-chatvoice/config route; settings persist to ~/.dsh/chatvoice.json
• client (client/client.js): MutationObserver injects the mic button (composer toolbar) and speaker buttons (assistant reply rows); SpeechRecognition for input; speechSynthesis for read-aloud
• Everything comes from the browser: the plugin makes no network requests, spawns no subprocesses, and needs no API key

Known limitations

• Chrome's speech recognition goes through Google servers — on some networks it reports a network error → switch to Edge (Azure)
• Edge's online voices need network access; offline it falls back to the system's local voices
• Firefox / Safari don't support SpeechRecognition (the mic button is disabled with a hint; read-aloud still works)
• Recognition accuracy depends on the browser and your microphone, not the plugin

Roadmap (Phase 2)

• 🎙 Push-to-talk (hold Space to dictate, release to send — WeChat-style)
• 🔊 edge-tts voices (XiaoxiaoNeural, generated server-side + attachment-route playback)
• 🗣 Voice commands ("save", "continue", "stop" and other spoken triggers)
• 📼 Voice memos: recordings transcribed into session drafts
• 🧩 An agent-callable read-aloud tool (host registers read_aloud, so the model can speak during replies)

License

MIT © FuzzySoul