baisama-cloud/dsh-stt-input

中文 · English

dsh-stt-input

Speech-to-text voice input for the DeepSeek Harness (DSH) web GUI.

Click the 🎤 mic button beside the composer, speak, click again — the transcript is written straight into the input box. The STT engine and model are selectable in Settings → 语音输入 (Voice Input).

Features

• Two engines
  • Browser local — uses the browser's built-in Web Speech API (SpeechRecognition, Chrome/Edge). Zero configuration, no API key, live interim results appear in the input box as you speak.
  • API — records with MediaRecorder and transcribes through any OpenAI-compatible /v1/audio/transcriptions endpoint (OpenAI, Groq, custom).
• Selectable model — whisper-1 (OpenAI), whisper-large-v3, whisper-large-v3-turbo, distil-whisper-large-v3-en (Groq), or a custom model name.
• Configurable — service preset (OpenAI / Groq / custom), API base URL, API key, recognition language, and insert mode (append to / replace the existing draft).
• Live status — a pill under the composer shows recording time, transcribing state, and errors.
• Privacy — the API key is kept in page memory only and is never persisted or logged; the non-secret config survives reloads via localStorage.

Install

Build the tarball and install it into your DSH web profile (same flow as other dsh-* plugins):

pnpm pack
# copy dsh-stt-input-*.tgz into the web profile and add the dependency,
# e.g. under ~/.dsh/profiles/web: pnpm add ../path/to/dsh-stt-input-0.1.0.tgz
# then restart `dsh web`.

The plugin registers:

• a mic button in the composer tool row (conversation.input.left),
• a status pill under the composer (conversation.composer.dock),
• a settings page (settings.section → 语音输入).

Usage

1. Open Settings → 语音输入 and pick an engine.
   • Browser local: nothing else needed (Chrome/Edge).
   • API: choose a preset (OpenAI or Groq), pick the model, and paste your API key. Groq's whisper-large-v3 is currently free.
2. Click the 🎤 mic button in the composer to start, speak, and click again to stop. The transcript lands in the input box; press Enter to send.

The browser engine needs Chrome or Edge (Web Speech API). Firefox falls back to the API engine.

How it works

┌──────────┐  click 🎤        ┌───────────────┐
│  Client  │ ───────────────▶ │ MediaRecorder │  (api engine)
│ (browser)│                  │ SpeechRecog.  │  (browser engine)
└──────────┘                  └──────┬────────┘
      ▲                              ▼
      │ setDraft(text)         base64 audio (JSON)
      │                 POST /stt-input/transcribe
      │                              │
┌─────┴──────┐              ┌───────▼────────┐
│ input box  │ ◀────────────│  Host (Node)   │
└────────────┘   {ok,text}  │  fetch → /v1/   │
                            │  audio/transcr. │
                            └─────────────────┘

The client (lib/client.js) captures audio and posts base64 JSON to the host route /stt-input/transcribe. The host (lib/index.js) decodes the audio and uploads it as multipart/form-data to ${baseUrl}/v1/audio/transcriptions using Node ≥ 18 global fetch / FormData / Blob.

License

MIT