UniAudio 2.0

### TL;DR

UniAudio 2.0 is a unified audio language model designed to handle text, speech, sound, and music. It introduces ReasoningCodec, a discrete audio codec that factorizes audio into reasoning and reconstruction tokens, enhancing both understanding and generation tasks. The model is trained on 100 billion text tokens and 60 billion audio tokens, demonstrating strong performance across various audio tasks.

Key Insights & Metrics

Key Features

• ReasoningCodec for audio factorization
• Unified autoregressive architecture for text and audio
• Trained on extensive text and audio datasets
• Strong few-shot and zero-shot generalization
• Competitive performance across speech, sound, and music tasks

→ Related Releases

Discussion