FireRedAudio: 9B Unified Audio LLM Released

FireRedTeam has released FireRedAudio and FireRedTTS3 on Hugging Face and GitHub. Based on a 9B parameter LLM, FireRedAudio features a design that decouples the representations of the audio encoder (understanding) and the RedAE pathway (generation), performing ASR, audio understanding, zero-shot TTS, instruction-based TTS, and voice editing in a single model.

It enables accurate temporal grounding and reasoning for 1 hour of audio, demonstrating competitive performance on major benchmarks such as MMAU, MMSU, and Seed-TTS-Eval. FireRedTTS3 is a unified speech generation and editing model that utilizes semantically enriched speech representations.