FunASR

FunASR is an industrial-grade, open-source speech recognition toolkit developed by the ModelScope team at Alibaba DAMO Academy. It is built for production voice workloads where accuracy, speed, and language coverage matter: the toolkit delivers up to 170x realtime transcription speed, supports more than 50 languages with particularly strong Chinese performance, and goes well beyond plain text with speaker diarization (telling who spoke when), emotion recognition, punctuation restoration, and streaming capabilities for live audio. What sets FunASR apart from typical ASR demos is its end-to-end production orientation. It ships pretrained models that are ready to deploy, offers an OpenAI-compatible API so it can drop into existing stacks, and provides both local and cloud deployment paths. This makes it suitable for real applications — call-center transcription, meeting captioning, voice assistants, media subtitling, and compliance monitoring — rather than just experimentation. The toolkit is part of the broader ModelScope model ecosystem, which gives it access to a wide library of continuously improved speech models. FunASR has attracted a large community of developers (tens of thousands of GitHub stars), especially among teams building Chinese-language voice products who need an open, self-hostable alternative to proprietary speech APIs. Because the models and runtime are open, teams can run everything on their own infrastructure for data sovereignty and predictable cost. It is aimed at developers and engineering teams who need accurate, fast, multilingual speech recognition they can self-host and integrate into products — particularly those with heavy Chinese-language requirements.