openclaw-whisper-voice
# OpenClaw Whisper Voice
Use this skill to make local Whisper transcription dependable on the OpenClaw Gateway host.
## Install on the host
Run:
```bash
{baseDir}/scripts/install_local_whisper.sh
```
The installer:
- installs Python packages into `~/.local`
- installs a CPU-safe PyTorch build
- installs `openai-whisper`
- installs `imageio-ffmpeg`
- creates stable `~/.local/bin/whisper` and `~/.local/bin/ffmpeg` launchers
## Transcribe a file manually
Use the wrapper instead of raw `whisper` when reliability matters:
```bash
{baseDir}/scripts/transcribe.sh /path/to/audio.ogg
{baseDir}/scripts/transcribe.sh /path/to/audio.m4a --model tiny --stdout-only
{baseDir}/scripts/transcribe.sh /path/to/audio.mp3 --task translate --format srt
```
## Configure inbound WhatsApp and Telegram voice notes
Patch OpenClaw config so inbound audio uses the wrapper:
```json5
{
tools: {
media: {
audio: {
enabled: true,
maxBytes: 20971520,
timeoutSeconds: 120,
models: [
{
type: "cli",
command: "{baseDir}/scripts/transcribe.sh",
args: ["{{MediaPath}}", "--model", "base", "--stdout-only"],
timeoutSeconds: 120
}
]
}
}
}
}
```
## Model choices
- `tiny`: fastest, weakest accuracy
- `base`: best default for chat voice notes
- `small` or larger: better accuracy, heavier CPU and RAM use
## Output rules
- Use `--stdout-only` for `tools.media.audio` so stdout is only transcript text.
- Use `--format txt|srt|vtt|json` for standalone file transcription.
- First model download goes into `~/.cache/whisper`.
标签
skill
ai
