Tencent open-sources Covo-Audio for full-duplex voice chat

Tencent just open-sourced Covo-Audio-Chat, a 7B end-to-end audio language model. Instead of cobbling together separate speech recognition, text LLM, and TTS pipelines, it runs everything inside a single architecture that takes continuous audio in and spits audio back out.

The project delivers native, low-latency full-duplex conversations. That means the model can listen while it speaks, allowing you to interrupt naturally mid-sentence just like talking to a human.

What Covo-Audio does

Most voice bots still operate like walkie-talkies where you speak, wait for silence detection, and then wait for audio synthesis.

Covo-Audio solves this by unifying continuous acoustic features, discrete speech tokens, and natural language text into a single interleaved sequence. This keeps prosodic nuances like cadence and emotional inflection intact without breaking semantic reasoning.

Another clever engineering choice is intelligence-speaker decoupling. Voice models often suffer when voice timbre is baked into the reasoning layers. Tencent trained the model across multiple speakers to separate the voice identity from the conversational brain, making it possible to transfer high-quality voices across different contexts without degrading IQ.

According to Tencent's technical report, the 7B model hits state-of-the-art or competitive numbers across spoken dialogue, speech comprehension, audio understanding, and full-duplex interaction among models in its weight class.

How to run it

Getting started requires Python 3.11 or higher.

You can clone the codebase from Tencent's GitHub repo and pull the model weights directly from the tencent/Covo-Audio-Chat Hugging Face repository into your local folder.

Tencent includes a one-click inference bash script called example.sh. Swap in your own input audio files in example.py, update the local paths, and you can test the interaction pipeline immediately.

Keep in mind that the release is currently licensed for research and experimental purposes only. Like any generative voice system, hallucination and edge-case handling still need human review.

Why it matters

Full-duplex end-to-end audio is clearly where conversational AI is heading, but most accessible tools are still clunky multi-hop pipelines. Open-sourcing a native 7B model lowers the bar significantly for developers building hardware companions, hands-free assistants, or real-time calling agents.

The 7B footprint strikes a sensible balance between hardware cost and conversational coherence. That said, real-world conversational latency and background noise handling will be the true tests before anyone bets their production stack on it.