FunASR is an open-source speech recognition toolkit covering training, inference, and production serving. Maintained within the ModelScope ecosystem, it has earned more than 20,700 stars on GitHub, is actively developed with updates in October 2026, and is available under the MIT license.
FunASR bundles the complete pipeline needed to build real speech products: voice activity detection, end-to-end speech recognition, punctuation restoration, speaker diarization, and streaming recognition for live audio. It ships with strong pretrained models such as Paraformer, with particular strength in Chinese and multilingual recognition. For deployment, it offers an OpenAI-compatible API and MCP server options, so recognized speech can plug straight into LLM applications and agent workflows without custom glue code.
FunASR is built to carry speech projects from experiment to production. Researchers can train and fine-tune models on their own data, while product teams can pick from several serving runtimes optimized for CPU or GPU deployment and expose recognition through standard APIs. Its models have been downloaded heavily through the ModelScope hub, and the toolkit is a common default for Mandarin speech tasks, where its pretrained models are known for strong accuracy on real-world audio with noise and accents. Regular releases add new models, including large speech foundation models, keeping the toolkit current as the field moves.
FunASR is completely free and open source under the MIT license, including its pretrained models, making it a cost-effective alternative to commercial speech APIs.
FunASR is for developers building transcription, meeting-notes, call-center analytics, and voice-assistant products — especially where Chinese or multilingual speech, self-hosting, and low running costs matter.