AAITopic

← Back to today's topics

Verified · Jul 6, 2026

HF + Cerebras 7/1 drop: an open-source real-time voice AI stack with Gemma 4 31B + Qwen3TTS + Parakeet

2 sources

Hugging Face's July 1, 2026 blog post describes a cascaded speech-to-speech pipeline: Nvidia Parakeet for ASR, Google DeepMind Gemma 4 31B VLM running on Cerebras for inference, and Alibaba Qwen3TTS for text-to-speech. The primary deployment target is the Reachy Mini robot platform (9,000+ in the wild per article; a community comment says 'crossed 10k'). The open-source library huggingface/speech-to-speech (Apache-2.0, default stack: Silero VAD v5 + Parakeet TDT + OpenAI-compatible LLM + Qwen3-TTS, latest v0.2.10 dated 2026-06-11) powers it. Live demo at hf.co/spaces/smolagents/hf-realtime-voice (218 running). The article does NOT publish latency numbers (ms / P95) — only qualitative comparisons.

Why now

On 7/1 this entry wires up six open / semi-open components — HF platform + Cerebras hardware + Gemma 4 model + Qwen3TTS + Parakeet + the speech-to-speech library — into a single public demo for the first time. Creators can use it as a 'reference blueprint' for open-source real-time voice AI without having to fabricate latency numbers.

Why it is worth publishing

The pipeline is open-source, reproducible, and the creators can break it down as a real example (no fabrication required). Latency numbers are absent, so the citation boundary is clear.

Evidence basis

Open-source real-time voice AI, Reachy Mini, Cerebras inference speed — each is a high-search-volume topic, and this post stitches the six pieces together into one ready-made architecture diagram.

“I wired six open components into a real-time voice AI pipeline — already on 9,000+ Reachy Minis.”

Angle

Frame this as a 'reference blueprint of six open / semi-open components wired together into a real-time voice AI pipeline', not as a new model release.

Format

Long-form explainer

Demo idea

Draw a 6-row architecture diagram: Nvidia Parakeet (ASR) → Google DeepMind Gemma 4 31B VLM (running on Cerebras inference) → Alibaba Qwen3TTS (TTS); add a row for the HF speech-to-speech library (Apache-2.0, v0.2.10, 2026-06-11, default stack Silero VAD v5 + Parakeet TDT + OpenAI-compatible LLM + Qwen3-TTS). Add a side row showing Reachy Mini deployment count (9,000+) and live demo instance count (218).

Platform notes

The article does NOT publish latency numbers (ms / P95). Creators should pin the link, quote only what the article states (cascaded pipeline components, open-source code, Reachy Mini deployment scale), and avoid inventing latency, throughput, or quality numbers. '9,000+ in the wild' is in the article; '10k+' is in a user comment — keep them separate.

Usable claims

  • Hugging Face's July 1, 2026 blog post describes a cascaded speech-to-speech pipeline that pairs Nvidia's Parakeet (ASR), Google DeepMind's Gemma 4 31B VLM running on Cerebras inference, and Alibaba's Qwen3TTS for text-to-speech, targeting real-time voice AI for applications including the Reachy Mini robot platform.
  • Hugging Face's July 1, 2026 blog post says the Reachy Mini robot platform is '9,000+ in the wild' (one community comment dated July 1 says the count 'crossed 10k'). The article frames these robots as the primary deployment target of the cascaded voice pipeline.

Evidence pipeline

From the news

Breakdown

This breakdown unpacks the 7/1 HF blog's cascaded speech-to-speech pipeline into six components: Nvidia Parakeet (ASR) → Google DeepMind Gemma 4 31B VLM (running on Cerebras) → Alibaba Qwen3TTS (TTS), plus the HF open-source library huggingface/speech-to-speech (Apache-2.0, v0.2.10, 2026-06-11, default stack: Silero VAD v5 + Parakeet TDT + OpenAI-compatible LLM + Qwen3-TTS). Reachy Mini is the primary deployment target (article says 9,000+ in the wild; live demo has 218 running instances). The article does NOT publish specific latency numbers (ms / P95) — creators should pin the link and avoid inventing them.

Sources

Risks

  • Pin a link to the HF blog post; quote only what the article states (cascaded pipeline components, open-source code, Reachy Mini deployment scale); do not invent latency, throughput, or quality numbers.

Demo ideas

  • Draw a 6-row architecture diagram labeling each component's vendor and role
  • Side-by-side Reachy Mini deployment count + live demo instance count (9,000+ / 218), showing 'real hardware + live demo' in parallel
  • Compare this 7/1 HF cascaded pipeline against OpenAI Realtime API's single-model end-to-end approach — two architectures for real-time voice