Tencent AI Open Sources Covo-Audio: A 7B Speech Language Model and Inference Pipeline for Real-Time Audio Conversations and Reasoning

Covo-Audio was released as open source released as an open-source 7B-parameter end-to-end large audio language model.

Audio Encoder uses Whisper-large-v3 as the primary encoder uses Whisper-large-v3 and operates at a frame rate of 50 Hz.

Audio Adapter downsamples audio frames from 50 Hz to 6.25 Hz employs three downsampling modules combining linear and convolution layers to reduce frame rate to 6.25 Hz.

LLM Backbone uses Qwen2.5-7B-Base as the backbone is based on Qwen2.5-7B-Base and adapted to process interleaved acoustic features and text tokens.

Speech Tokenizer generates discrete audio tokens at 25 Hz with a 16,384 codebook is based on WavLM-large with a codebook size of 16,384 and produces discrete audio tokens at 25 Hz.