Fun-ASR
A family of end-to-end speech recognition models from Tongyi Lab, covering many languages and Chinese regional dialects and accents, deployable through Transformers, FunASR, vLLM or llama.cpp.
Fun-ASR is a family of end-to-end speech recognition models from Tongyi Lab, distributed with native Hugging Face Transformers support for its Nano checkpoint alongside separate FunASR, vLLM and llama.cpp deployment paths. Fun-ASR-Nano is an 800-million-parameter checkpoint trained on tens of millions of hours of speech, recognizing Chinese, English and Japanese, including seven Chinese dialects and more than twenty regional accents across both languages, plus lyric and rap-speech recognition.
A second checkpoint, Fun-ASR-MLT-Nano, is also an 800-million-parameter model but trained for broader multilingual coverage, spanning thirty-one languages including Cantonese, Korean, Vietnamese, Thai, Arabic and several European languages. Both checkpoints integrate with FunASR for inference and serving, and the repository provides runnable examples covering quickstart and direct inference, speaker diarization, vLLM batch inference and a streaming SDK, plus online demo spaces on ModelScope and Hugging Face. It is written in C and licensed under Apache 2.0.