Most voice AI pipelines are cascaded: ASR transcribes, an LLM reasons, TTS speaks. Covo-Audio collapses all three into a single 7B model that takes audio in and produces audio out, built on Qwen2.5-7B with a Whisper-large-v3 encoder.
The full-duplex variant (Covo-Audio-Chat-FD) is the interesting part. It handles simultaneous listening and speaking using three special tokens baked into the architecture: THINK (listening, not yet responding), SHIFT (switching to speaking turn), and BREAK (user interrupted, stop speaking immediately). Each audio chunk is 0.16 seconds. The model and user streams are interleaved at a 1:4 ratio.
The intelligence-speaker decoupling technique is also notable: dialogue intelligence and voice rendering are trained separately, so you can customize the voice using minimal TTS data without retraining the reasoning layer.
One caveat worth noting: a GitHub issue filed the day of release points out that the full-duplex inference code is not yet in the public repo. The paper shows strong results but it’s unclear whether those are reproducible from the current open-source release.