audiosronnx-frcrn

ONNX export of FRCRN_SE_16K from ClearerVoice-Studio (Alibaba, Apache-2.0), for the frcrn denoise engine in 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

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.