Neura Market
FunASR logo

FunASR

Free

Industrial-grade speech recognition toolkit: 170x realtime, 50+ languages, speaker diarization, emotion detection, streaming, and OpenAI-compatible API.

FreeFree tier
Inputs: audioOutputs: text
Type
Open Source
Visit Website GitHub (16.2k stars)

About FunASR

FunASR is an industrial-grade, open-source end-to-end speech understanding toolkit that provides a unified Python interface for automatic speech recognition (ASR), voice activity detection (VAD), punctuation restoration, speaker diarization, emotion detection, and audio event recognition. It supports 50+ languages, achieves 170x realtime processing, and offers an OpenAI-compatible API for easy integration into agents, apps, and batch pipelines. The toolkit includes multiple models such as Paraformer, SenseVoice, and Fun-ASR-Nano, and features a vLLM inference engine for 2-3x faster LLM decoding. It is designed for production use with deployment options including Python API, Docker Compose, Kubernetes, WebSocket, vLLM, MCP, and Triton. FunASR also supports fine-tuning, benchmarks, and agent integration, making it suitable for private speech APIs, streaming transcription, subtitle generation, and research.

Key Features

End-to-end ASR with VAD segmentation and punctuation restoration
Speaker diarization and emotion detection in a single pipeline
Supports 50+ languages with models like Fun-ASR-Nano (31 languages) and Qwen3-ASR (52 languages with language detection)
170x realtime performance on GPU
OpenAI-compatible API for easy integration with existing applications
vLLM inference engine for 2-3x faster LLM decoding with tensor parallel batch inference
Agent infrastructure with MCP tools and voice-input examples
Deployment options: Python API, Docker Compose, Kubernetes, WebSocket, vLLM, MCP, batch, subtitles, Triton
Long-form audio processing with a single API call
Fine-tuning support for Paraformer, SenseVoice, and Fun-ASR-Nano on custom data

Pros & Cons

Pros
  • Production-ready and open source under the MIT license with 16,000+ GitHub stars
  • Comprehensive speech understanding pipeline combining ASR, VAD, punctuation, diarization, emotion, and event detection
  • High performance: 170x realtime and 2-3x faster LLM decoding with vLLM
  • Easy integration via OpenAI-compatible API, MCP, and low-code workflow nodes
  • Active development with regular updates, benchmarks, and community contributions
  • Fine-tuning support allows customization for specific domains and languages
  • Multiple deployment options suitable for cloud, on-premises, or edge
Cons
  • Requires Python and GPU for optimal real-time performance; CPU inference is slower
  • Steeper learning curve compared to turnkey cloud ASR services
  • Documentation and tutorials are still maturing (tracking latest README)
  • Some models (e.g., Paraformer) may be optimized primarily for Mandarin Chinese, though multilingual models are available

Best For

Private speech API deployment without sending audio to cloud providersVoice input agents and AI tool integration via OpenAI-compatible endpoints or MCPStreaming and batch transcription of meetings, calls, or lecturesSubtitle generation for videos and multilingual contentSpeaker-attributed transcription for meetings and interviewsBenchmarking and research comparison against Whisper and cloud ASRFine-tuning models on domain-specific audio data

FAQ

What languages does FunASR support?
FunASR supports 50+ languages. Models like Fun-ASR-Nano cover 31 languages, and Qwen3-ASR covers 52 languages with automatic language detection.
Is FunASR free to use?
Yes, FunASR is open source under the MIT license and completely free to use, with no cloud API costs.
Can I run FunASR without a GPU?
Yes, FunASR can run on CPU, but GPU is recommended for 170x realtime performance. The website provides benchmark results for both GPU and CPU.
Does FunASR support streaming ASR?
Yes, FunASR supports streaming ASR with VAD segmentation and can be deployed with a WebSocket service for real-time applications.
Can I fine-tune FunASR models on my own data?
Yes, the training guide provides instructions to fine-tune Paraformer, SenseVoice, and Fun-ASR-Nano models on JSONL-formatted custom data.
How do I integrate FunASR with agents or existing tools?
FunASR offers an OpenAI-compatible API server, MCP tools, and voice-input examples to connect local ASR with AI tools and agents.