Voice Input with Whisper & FFmpeg

Voice: the highest-bandwidth input you have

Typing is slow. Talking to your agent is two to three times faster, and on a phone it is the only humane option. We use Whisper for speech-to-text and FFmpeg for the audio pipeline.

Two flavors of Whisper

• whisper.cpp — pure C++, runs on CPU, 100% local, free
• OpenAI Whisper API — cloud, fast, paid, slightly more accurate on noisy audio

For privacy and zero ongoing cost, start with whisper.cpp. We will fall back to the API for hard cases.

Install whisper.cpp

git clone https://github.com/ggerganov/whisper.cpp ~/Code/whisper
cd ~/Code/whisper && make
./models/download-ggml-model.sh base.en

Wire it into OpenClaw

Add a Skill or Tool entry that:

1. Records audio from the mic with FFmpeg into a temp .wav
2. Pipes the file through whisper.cpp to produce text
3. Sends the text into OpenClaw as if it were a typed message

Example FFmpeg recorder line (macOS):

ffmpeg -f avfoundation -i ":0" -t 30 -ar 16000 -ac 1 /tmp/voice.wav

When to choose which model

• tiny.en — keyword grade, fastest
• base.en — daily driver, ~1× realtime on a modern laptop
• small.en / medium.en — meeting transcription, slower but punchy

Try it

Record yourself dictating a one-paragraph task. Confirm OpenClaw receives the transcribed text and acts on it.