Google, Hugging Face, and Cerebras open-source ultra-fast voice AI

Signals Inbox·July 20, 2026·Voice AI

Google, Hugging Face, and Cerebras have released an open speech-to-speech stack powered by Gemma 4 31B. The headline is not simply that Gemma can talk. Cerebras says the model runs at 1,851 output tokens per second, roughly 35 times a typical GPU endpoint, making a capable open model fast enough for live voice conversations.

Q1What actually launched?

Hugging Face and Cerebras released a fully open, modular speech-to-speech pipeline using Google DeepMind’s Gemma 4 31B as its brain. The stack turns speech into text, sends it through Gemma, and converts the answer back into speech. Developers can inspect, replace, and modify each component.

Q2How fast is it?

Cerebras says Gemma 4 31B runs at 1,851 output tokens per second on its hardware. That is about 35 times faster than a typical GPU endpoint. It returns its first token, including reasoning time, in roughly 1.5 seconds. The full spoken reply also needs speech recognition and voice generation, but the slowest thinking step is now much faster.

Q3Why does voice AI need that much speed?

Because normal conversation has very little patience for silence. A two or three-second pause after every sentence makes an assistant feel mechanical. Voice systems also make several model calls for tools, memory, safety checks, and follow-up questions. Small delays quickly stack up. Faster and more predictable inference makes interruption, back-and-forth conversation, and live robot control feel possible.

Q4Is this one native voice model?

No. It is a cascaded system. Nvidia’s Parakeet handles speech recognition, Gemma 4 produces the answer on Cerebras, and Alibaba’s Qwen3-TTS creates the spoken response. Native audio models can sometimes sound smoother because they process speech directly. The advantage here is control: developers can swap any layer without replacing the entire product.

Q5How does it compare with closed voice APIs?

OpenAI, Google, and other large platforms already offer fast voice systems that combine listening, reasoning, and speaking. Their advantage is tighter integration and polished voices. This stack attacks their other advantage: speed. Developers can now use an open 31B model without accepting the slow replies normally associated with self-hosted or open-model voice apps.

Q6Is anyone already using this architecture?

Yes. Hugging Face says the same speech-to-speech pipeline already powers more than 9,000 Reachy Mini robots in the wild. That makes this more than a lab demo. Robots are an especially hard test because a delayed reply breaks the illusion that the machine is listening and reacting in real time.

Q7So what is the real signal?

Open voice AI is starting to lose its speed penalty. The important number is not 31 billion parameters but 1,851 tokens per second. If open models can match the responsiveness users expect from closed APIs, voice developers gain more control over data, costs, models, and deployment. The next test is whether the full pipeline stays fast and reliable under real production traffic.