FunASR
An open-source speech-to-text toolkit supports offline files and live transcription on your hardware, with speaker labels and API serving.
FunASR is a speech-to-text toolkit for developers building transcription services and voice applications. It supports offline recordings and live speech on local hardware or self-hosted servers, with capabilities that depend on the model and runtime.
- Multilingual transcription: Fun-ASR-Nano covers Chinese, English, Japanese and Chinese dialects and accents; Fun-ASR-MLT-Nano provides broader coverage, with Qwen3-ASR and Whisper also available.
- Live speech: Paraformer-zh-streaming handles chunked transcription through streaming interfaces, separately from models intended for recorded files.
- Speaker labels and timing: Combine SenseVoiceSmall, FSMN-VAD and CAM++ for speaker-aware segments, or use the OpenMOSS MOSS-Transcribe-Diarize adapter for joint text, timestamps and speakers in recorded audio; labels distinguish speakers within a recording rather than identify known people.
- Audio tags: SenseVoiceSmall adds emotion and audio-event tags to transcription in Chinese, English, Japanese, Korean and Cantonese.
- Application connections: Serve transcription through an OpenAI-compatible API or MCP for tools including Claude and Cursor.
CPU and NVIDIA GPU execution are supported, with vLLM available for Nano batch processing. The llama.cpp runtime uses compatible GGUF weights on CPU and edge devices without Python at runtime, with Linux and Windows GPU packages for supported paths. Colab provides a separate cloud option.
The toolkit is open source under MIT. Model weights have separate licenses, and features available through one model or adapter don't necessarily work through every serving backend.