FunASR (Paraformer)
Industrial-grade ASR toolkit by Alibaba with Paraformer non-autoregressive models.
About
FunASR is an industrial-grade speech recognition toolkit from Alibaba, published through ModelScope, that covers offline, streaming, and edge deployment rather than betting on a single model. Its catalog includes Paraformer, a 220M-parameter non-autoregressive recognizer for Chinese and English with timestamp output, SenseVoiceSmall, which adds emotion recognition and audio event detection across five languages, and larger Fun-ASR and multilingual variants reaching dozens of languages. Around the recognizers sit voice activity detection, punctuation restoration, and speaker diarization modules that combine into complete transcription pipelines. The models are efficient enough that CPU inference can outpace Whisper on GPU, vLLM acceleration is reported at up to 340x realtime, and GGUF builds run through llama.cpp on edge devices without Python. Deployment options include Docker, WebSocket streaming servers, and OpenAI-compatible APIs. The code is MIT licensed with model weights licensed separately, and it is widely used in production Chinese ASR systems.
Should you use FunASR (Paraformer)?
Pick it when
Pick FunASR when you need a complete Chinese-first transcription pipeline, recognizer plus VAD, punctuation, and diarization, deployed as Docker, WebSocket streaming, or an OpenAI-compatible API, where CPU throughput matters.
Look elsewhere when
Skip it if licensing must be simple, since the code is MIT but model weights carry separate licenses, or if your audio is mostly English or broadly multilingual, where Whisper-based tools have the larger ecosystem.
Alternatives to FunASR (Paraformer)
- SenseVoice
The SenseVoiceSmall model on its own, with emotion and event tags and ONNX or GGUF export, when you want one model rather than FunASR's VAD, punctuation, and diarization pipeline.
- WeNet
Production Conformer toolkit with libtorch, ONNX, and TensorRT runtimes, better if you plan to train and compile your own models.
- WhisperX
Whisper pipeline with VAD batching, word alignment, and diarization, a stronger fit for English and multilingual meetings than Chinese-first FunASR.
- Qwen3-ASR
Apache 2.0 weights covering 30 languages and 22 Chinese dialects, simpler licensing than FunASR's separately licensed models.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- MIT
- Minimum VRAM
- 4 GB
- Added
- Apr 3, 2026