FireRedAudio

FireRedAudio is a 9-billion-parameter general-purpose audio language model from the FireRedTeam (Xiaohongshu), released in August 2026 with code and weights under Apache 2.0. It unifies speech recognition, audio understanding, zero-shot text-to-speech, instruct TTS, and speech editing in a single trainable autoregressive LLM instead of a cascade pipeline.

Its signature design is decoupled continuous representations: an Audio Encoder handles input for understanding, while a RedAE-based pathway handles generation. A single model covers ASR, hour-long audio understanding, zero-shot and instruct TTS, and semantic plus acoustic speech editing, with temporal grounding reported at second-level accuracy.

The release also includes FireRedTTS3, a speech synthesis and editing framework with Base and Instruct variants. Coverage is strongest in Chinese and English, and the project page cautions the research-purpose framing common to such releases, so production voice-cloning users should review consent obligations.