Paraformer-zh

Non-autoregressive end-to-end speech recognition — 120x realtime on GPU, production-ready for Mandarin Chinese.

Paraformer is a non-autoregressive (NAR) ASR model that generates the entire output in parallel, achieving significant speedups over autoregressive models like Whisper while maintaining competitive accuracy.

Quick Start

from funasr import AutoModel

# Basic recognition
model = AutoModel(model="funasr/paraformer-zh", hub="hf", device="cuda")
result = model.generate(input="audio.wav")
print(result[0]["text"])

Full Pipeline (VAD + ASR + Punctuation + Speaker Diarization)

from funasr import AutoModel

model = AutoModel(
 model="funasr/paraformer-zh",
 hub="hf",
 vad_model="funasr/fsmn-vad",
 punc_model="funasr/ct-punc",
 spk_model="funasr/campplus",
 device="cuda",
)

result = model.generate(input="meeting.wav")
# Output includes timestamps, punctuation, and speaker labels
for sentence in result[0]["sentence_info"]:
 print(f"[Speaker {sentence['spk']}] {sentence['text']}")

Features

• 120x realtime on GPU (non-autoregressive parallel decoding)
• Chinese + English mixed recognition
• Built-in VAD (voice activity detection) for long audio
• Punctuation restoration with ct-punc model
• Speaker diarization with cam++ model
• Streaming and offline modes
• ONNX export supported

Model Details

Related Models

Links

• GitHub: FunASR
• Docs: modelscope.github.io/FunASR
• Paper: Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition

Citation

@inproceedings{gao2022paraformer,
 title={Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition},
 author={Gao, Zhifu and Zhang, Shiliang and McLoughlin, Ian and Yan, Zhijie},
 booktitle={INTERSPEECH},
 year={2022}
}