Voice Review

README

# openclaw-voice-review

OpenClaw plugin that intercepts incoming voice messages, transcribes (and optionally translates to English) them using OpenRouter's Whisper endpoint, then presents the text for user review before it reaches the LLM.

## How it works

1. Voice message arrives
2. Plugin intercepts via `inbound_claim` hook (before LLM sees it)
3. Audio is base64-encoded and sent to OpenRouter `/api/v1/audio/transcriptions`
4. Transcription/translation is shown as a monospaced preview with a confirmation hint
5. User replies "ok" to submit (or edits and sends manually)
6. The text flows to the LLM as a normal user message

## Configuration

Add to your `openclaw.json`:

```json5
{
  plugins: {
    entries: {
      "voice-review": {
        enabled: true,
        config: {
          mode: "translate_to_en",  // "transcribe" or "translate_to_en"
          model: "openai/whisper-large-v3-turbo",
          confirmMessage: "Reply ok to submit, or edit and send.",
          promptText: ""  // optional Whisper prompt
        }
      }
    }
  }
}
```

### Settings

| Setting | Default | Description |
|---|---|---|
| `mode` | `translate_to_en` | `transcribe` = original language; `translate_to_en` = transcribe + translate to English |
| `model` | `openai/whisper-large-v3-turbo` | OpenRouter Whisper model |
| `confirmMessage` | `Reply ok to submit, or edit and send.` | Hint shown below preview |
| `promptText` | (empty) | Optional Whisper context prompt |

### Requirements

- `OPENROUTER_API_KEY` environment variable must be set
- OpenRouter account with access to Whisper models

### Audio scope

The plugin only intercepts audio that would normally be transcribed by OpenClaw's built-in audio processing. Make sure `tools.media.audio` scope allows the chat types where you want this to work.

## Installation

```bash
# From source
git clone https://github.com/BorClaw/openclaw-voice-review.git
# Add to openclaw.json plugins.entries with localPath
```

## Cost

Whisper Large V3 Turbo via OpenRouter: ~$0.0001 per second of audio. A 30-second voice message costs less than $0.01.

## License

MIT