OpenClaw Voice Assistant

Yes. Whisper large-v3 (speech-to-text) and Kokoro (text-to-speech) both run directly on the ClawBox's NVIDIA Jetson GPU. Your spoken audio is transcribed on-device and the spoken reply is synthesized on-device. No audio is ever uploaded to a cloud speech API — voice works even with the network cable unplugged, making it ideal for private offices or locations with poor connectivity.

Whisper is OpenAI's open-source speech-to-text model; ClawBox runs the large-v3 variant locally with automatic language detection across 90+ languages and built-in Voice Activity Detection. Kokoro is a lightweight open-source text-to-speech model that produces natural-sounding speech with multiple voice options. Together they form a complete local voice pipeline: you speak, Whisper transcribes, the AI responds, and Kokoro speaks the answer back.

Completely private. Unlike Siri (Apple), Alexa (Amazon), or Google Assistant — which stream your voice to company servers — ClawBox transcribes and synthesizes everything locally. Your recordings, transcripts, and spoken replies never leave the device unless you explicitly route a request to an external API. OpenClaw is open source, so you can verify there is no telemetry or hidden voice upload.

Whisper large-v3 transcribes speech at near-realtime speed on the 67 TOPS Jetson GPU, and Kokoro typically generates the spoken reply within a couple of seconds. Total round-trip latency depends mostly on how long the language model takes to think, not on the voice layer. Because nothing travels to the cloud, there is no network round-trip adding delay, and performance stays consistent regardless of internet quality.

Yes. You can send voice messages to your AI bot on Telegram from anywhere and receive voice or text responses, all processed locally on the ClawBox at home. Whisper automatically detects the spoken language across 90+ languages, so you can speak Japanese, switch to German mid-conversation, and the assistant keeps up — with no per-minute cloud speech costs.