---
license: apache-2.0
library_name: audiosronnx
pipeline_tag: audio-to-audio
tags:
  - speech-enhancement
  - denoising
  - noise-suppression
  - onnx
---

# audiosronnx-frcrn

ONNX export of **FRCRN_SE_16K** from
[ClearerVoice-Studio](https://github.com/modelscope/ClearerVoice-Studio) (Alibaba,
Apache-2.0), for the `frcrn` denoise engine in
[audiosronnx](https://github.com/TigreGotico/audiosronnx).

FRCRN (*Frequency Recurrent CRN*) placed 2nd in the 2022 DNS Challenge and reports
**PESQ 3.23 / STOI 0.95 / SI-SDR 19.22** on VoiceBank+DEMAND.

The graph is waveform-to-waveform at 16 kHz with a dynamic length axis:

```
noisy[1, T]  ->  enhanced[1, T]
```

Its STFT/ISTFT are `ConvSTFT`/`ConviSTFT` — Conv1d layers with Fourier kernels, not
`torch.stft` — so the whole two-UNet complex-mask pipeline exports as one static graph
and inference is onnxruntime-only.

| File | Rate | Size |
|------|------|------|
| `frcrn.onnx` | 16 kHz | 57.5 MB |

Validated against the upstream ClearerVoice pipeline end-to-end: **correlation
0.99999994, 57.6 dB SNR** (residual is fp32 rounding).

## Usage

```python
from audiosronnx import load_denoise

dn = load_denoise("frcrn")
clean, rate = dn.denoise("noisy.wav")   # -> (float32 mono, 16000)
```

## License

Apache-2.0, inherited from ClearerVoice-Studio.