FireRedASR 2

FireRedASR 2 is the second-generation automatic speech recognition model from the FireRedTeam, building on the original FireRedASR released in early 2025. It targets high-accuracy Chinese transcription, covering Mandarin, more than 20 dialects and accents, English, code-switching, and even singing lyrics. The team reports a 2.89 percent average character error rate on Mandarin benchmarks and 11.55 percent across Chinese dialect sets, outperforming several contemporary open ASR systems on those tests.

The AED variant adds word-level timestamps and confidence scores, and the broader FireRedASR2S release bundles the recognizer with voice activity detection, language identification, and punctuation modules. Release date is era-level (early 2025) since sources document the paper and model drops across January and February 2025. The project is open source with weights on Hugging Face.