Hugging Face Voice-Agent Stack Adds Native-Audio Path With Gemma 4 12B

Hugging Face's open-source speech-to-speech project now documents a local voice-agent path that sends spoken audio directly into a multimodal language model, replacing the usual speech-to-text hop with VAD -> Gemma 4 12B -> TTS. The change points to a simpler architecture for low-latency voice agents, though the implementation is still bounded by local hardware, experimental audio support and model-specific quirks.