Modular. Open. Local-first. And it runs on a MacBook.
On July 1, 2026, Hugging Face and Cerebras Systems shipped something the open-source community has been waiting for: a fully open, real-time speech-to-speech AI pipeline that goes head-to-head with OpenAI's Realtime API — without the vendor lock-in.
What's in the stack?
The pipeline chains together three best-in-class open models:
- Nvidia Parakeet — for automatic speech recognition (ASR), turning raw audio into text with sub-100ms latency.
- Google DeepMind's Gemma 4 31B — running on Cerebras wafer-scale hardware for lightning-fast reasoning and response generation.
- Alibaba's Qwen3TTS — a text-to-speech model that delivers natural, expressive voice output.
The glue? Hugging Face's 🤗 Transformers and a lightweight orchestration layer that makes the whole thing work as a single voice-pipeline call. Total round-trip latency? Under 500ms — competitive with proprietary offerings.
Why this matters
OpenAI's Realtime API costs $0.06 per minute of audio input and $0.24 per minute of audio output. This stack runs on your own hardware (or Cerebras' cloud at a fraction of the cost), and every component is Apache 2.0 or MIT licensed.
It's already been deployed on 9,000+ Reachy Mini robots for zero-latency voice interaction — think warehouse assistants, reception kiosks, and elderly care companions that actually understand you.
The bigger picture
This isn't just a pipeline — it's a template. Each component (ASR, LLM, TTS) can be swapped for alternatives. Want to use Whisper instead of Parakeet? Swap it in. Prefer Llama 4 over Gemma? That works too.
The voice AI duopoly just got a credible open-source challenger. And it's running on your laptop right now.
Links: Hugging Face model hub | Cerebras inference | Demo on Reachy robots
Top comments (0)