A tiny Pixi-managed local Whisper setup for OpenClaw audio / voice-note transcription.
This is the simple working pattern Charles and CTBot ended up using:
- keep Telegram as the chat channel
- use a local Whisper CLI for transcription
- run Whisper inside a Pixi environment so
ffmpegis available onPATH
This repo is not a full OpenClaw install.
It is just the little helper environment that gives OpenClaw a reliable local speech-to-text command for Telegram voice notes and other audio files.
- Pixi
- an existing OpenClaw setup
- a machine where local transcription is acceptable
pixi installOptional smoke test:
pixi run whisper --helppixi.toml— the Pixi environmenttranscripts/— output directory for Whisper text filesexamples/openclaw-audio-config.json— example OpenClaw audio transcription config snippet
The included task writes transcript text files into transcripts/:
pixi run transcribe -- your-audio-file.oggOr call Whisper directly:
pixi run whisper --model small --output_format txt --output_dir transcripts your-audio-file.oggThe key detail is to point OpenClaw at Pixi, not directly at a bare Whisper binary.
That keeps the Pixi-managed environment intact, including ffmpeg.
See:
examples/openclaw-audio-config.json
The important bit is using a command shaped like:
{
"type": "cli",
"command": "pixi",
"args": [
"run",
"-m",
"/Users/YOUR_USERNAME/voice-transcription",
"-x",
"whisper",
"--model",
"small",
"--output_format",
"txt",
"--output_dir",
"/Users/YOUR_USERNAME/voice-transcription/transcripts",
"{{MediaPath}}"
]
}Replace YOUR_USERNAME with the actual user running OpenClaw.
- This is intentionally small and practical.
- It is aimed at the “make local voice note transcription actually work” problem, not at being a general speech stack.
- Model choice is currently
smallbecause that was the working compromise in the original setup.