---
license: other
license_name: model-license
license_link: https://github.com/alibaba-damo-academy/FunASR
language:
- en
pipeline_tag: automatic-speech-recognition
tags:
- FunASR
- Paraformer
- ASR
- automatic-speech-recognition
- non-autoregressive
- speech-recognition
library_name: funasr
---

<div align="center">

# Paraformer-en

**A 220M offline, non-autoregressive English speech recognition model for 16 kHz audio.**

Powered by [FunASR](https://github.com/modelscope/FunASR). Follow the project or give it a star to receive new model, runtime, and deployment updates.

[FunASR](https://github.com/modelscope/FunASR) · [Fun-ASR](https://github.com/QwenAudio/Fun-ASR) · [SenseVoice](https://github.com/QwenAudio/SenseVoice) · [FunClip](https://github.com/modelscope/FunClip)

</div>

Paraformer predicts output tokens in parallel instead of decoding them one by one. This checkpoint is intended for high-throughput English transcription where complete utterances or files are available before decoding.

## Quick start

Install PyTorch and FunASR:

```bash
python -m pip install -U torch torchaudio funasr
```

For CUDA-specific PyTorch wheels, select the command for your platform in the [PyTorch installation guide](https://pytorch.org/get-started/locally/), then install `funasr`.

Run the audio example included in this repository:

```python
from funasr import AutoModel

model = AutoModel(
    model="funasr/paraformer-en",
    hub="hf",
    device="cpu",  # Change to "cuda:0" for NVIDIA GPU inference.
)

audio = f"{model.model_path}/example/asr_example.wav"
result = model.generate(input=audio)
print(result[0]["text"])
```

Using the full repository ID is recommended. It selects this English checkpoint directly and avoids ambiguity with similarly named Paraformer models.

## Long audio pipeline

Add voice activity detection to split long recordings and punctuation restoration to produce readable text:

```python
from funasr import AutoModel

model = AutoModel(
    model="funasr/paraformer-en",
    hub="hf",
    vad_model="funasr/fsmn-vad",
    punc_model="funasr/ct-punc",
    device="cuda:0",
)

result = model.generate(
    input="meeting.wav",
    batch_size_s=300,
)
print(result[0]["text"])
```

The pipeline accepts common audio file paths and 16 kHz waveform input. Resample other sample rates to 16 kHz for predictable results.

## OpenAI-compatible serving

Expose the model behind a local `/v1/audio/transcriptions` endpoint:

```bash
python -m pip install -U torch torchaudio funasr fastapi uvicorn python-multipart
funasr-server --model-path funasr/paraformer-en --hub hf --device cuda
```

```bash
curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@audio.wav \
  -F model=custom
```

`funasr-server` registers a checkpoint supplied through `--model-path` under the API model name `custom`.

See the [OpenAI-compatible API examples](https://github.com/modelscope/FunASR/tree/main/examples/openai_api), [deployment matrix](https://github.com/modelscope/FunASR/blob/main/docs/deployment_matrix.md), and [FunASR deployment hub](https://www.funasr.com/) for clients, containers, security guidance, and production topologies.

## Model details

| Property | Value |
| --- | --- |
| Architecture | Paraformer, non-autoregressive ASR |
| Parameters | 220M |
| Language | English |
| Input | 16 kHz speech audio |
| Mode | Offline / non-streaming |
| Framework | PyTorch through FunASR |
| Vocabulary | 10,020 tokens |

### Artifact provenance

This Hugging Face repository mirrors the PyTorch checkpoint published as [`iic/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020`](https://www.modelscope.cn/models/iic/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020) on ModelScope. It is not a TensorFlow 1.x Paraformer variant.

For reproducibility, the `model.pt` artifact in revision `d043d60434a65ae49b2b2462a09660049871b8e9` has:

- size: `887,289,223` bytes
- SHA-256: `e11203abf1f449b4c6515dee494c017ee27a4b716730378473d827ad3c7fe41c`

## Intended use and limitations

This model is suitable for offline English transcription, subtitle preparation, call or meeting processing, and batch speech pipelines. It is not a streaming checkpoint; use a streaming Paraformer model when partial low-latency results are required.

Recognition quality can degrade with heavy noise, overlapping speakers, strong accents, uncommon proper nouns, or domain-specific terminology. Validate accuracy on representative audio before production use, and review transcripts before using them in safety-critical or high-impact decisions.

## Related models

| Model | Purpose |
| --- | --- |
| [funasr/fsmn-vad](https://huggingface.co/funasr/fsmn-vad) | Voice activity detection for long audio |
| [funasr/ct-punc](https://huggingface.co/funasr/ct-punc) | Chinese and English punctuation restoration |
| [funasr/paraformer-zh](https://huggingface.co/funasr/paraformer-zh) | Offline Mandarin and English ASR |
| [funasr/paraformer-zh-streaming](https://huggingface.co/funasr/paraformer-zh-streaming) | Streaming Paraformer ASR |

## Resources

- [FunASR GitHub repository](https://github.com/modelscope/FunASR)
- [FunASR documentation](https://modelscope.github.io/FunASR/)
- [Paraformer paper](https://arxiv.org/abs/2206.08317)
- [Issue tracker](https://github.com/modelscope/FunASR/issues)

## Citation

```bibtex
@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}
}
```