funasr/Paraformer-large · Hugging Face

Introduce

Paraformer is an innovative non-autoregressive end-to-end speech recognition model that offers significant advantages over traditional autoregressive models. Unlike its counterparts, Paraformer can generate the target text for an entire sentence in parallel, making it ideal for parallel inference using GPUs. This capability leads to significant improvements in inference efficiency, which can reduce machine costs for speech recognition cloud services by almost 10 times. Furthermore, Paraformer can achieve the same performance as autoregressive models on industrial-scale data.

This repository demonstrates how to leverage Paraformer in conjunction with the funasr_onnx runtime. The underlying model is derived from FunASR, which was trained on a massive 60,000-hour Mandarin dataset. Notably, Paraformer's performance secured the top spot on the SpeechIO leaderboard, highlighting its exceptional capabilities in speech recognition.

We have relesed numerous industrial-grade models, including speech recognition, voice activity detection, punctuation restoration, speaker verification, speaker diarization, and timestamp prediction (force alignment). To learn more about these models, kindly refer to the documentation available on FunASR. If you are interested in leveraging advanced AI technology for your speech-related projects, we invite you to explore the possibilities offered by FunASR.

Install funasr_onnx

pip install -U funasr_onnx
# For the users in China, you could install with the command:
# pip install -U funasr_onnx -i https://mirror.sjtu.edu.cn/pypi/web/simple

Download the model

git clone https://huggingface.co/funasr/paraformer-large

Inference with runtime

Speech Recognition

Paraformer

from funasr_onnx import Paraformer

model_dir = "./paraformer-large"
model = Paraformer(model_dir, batch_size=1, quantize=True)

wav_path = ['./funasr/paraformer-large/asr_example.wav']

result = model(wav_path)
print(result)

model_dir: the model path, which contains model.onnx, config.yaml, am.mvn
batch_size: 1 (Default), the batch size duration inference
device_id: -1 (Default), infer on CPU. If you want to infer with GPU, set it to gpu_id (Please make sure that you have install the onnxruntime-gpu)
quantize: False (Default), load the model of model.onnx in model_dir. If set True, load the model of model_quant.onnx in model_dir
intra_op_num_threads: 4 (Default), sets the number of threads used for intraop parallelism on CPU

Input: wav formt file, support formats: str, np.ndarray, List[str]

Output: List[str]: recognition result

Performance benchmark

Please ref to benchmark

Citations

@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}
}