audiosronnx-mpsenet
ONNX exports of MP-SENet (Lu et al., MIT) for the
mpsenet denoise engine in
audiosronnx.
MP-SENet predicts the magnitude and phase spectra in parallel, rather than applying a magnitude mask and reusing the noisy phase. At 2.26 M parameters it is the smallest full spectral model in the library.
amp[1, 201, T], pha[1, 201, T] -> amp_g[1, 201, T], pha_g[1, 201, T]
The STFT, the 0.3 power-compression of magnitudes, and the utterance RMS normalisation all run in numpy inside audiosronnx, so inference is onnxruntime-only.
| File | Checkpoint | Rate | Size |
|---|---|---|---|
mpsenet_dns.onnx | g_best_dns (DNS Challenge) | 16 kHz | 9.7 MB |
mpsenet.onnx | g_best_vb (VoiceBank+DEMAND) | 16 kHz | 9.7 MB |
The two checkpoints are not interchangeable. Measured SNR gain on speech corrupted with broadband Gaussian noise, at 19 / 11 / 5 dB input SNR:
| Checkpoint | Gain |
|---|---|
dns | +4.8 / +8.8 / +11.7 dB |
vb | +1.5 / +2.0 / +4.1 dB |
dns is the default. vb is the checkpoint the published VoiceBank+DEMAND PESQ figures
come from, and is the one to use when reproducing those numbers — but it generalises poorly
to noise unlike its training set.
Export note
The model uses stock torch.nn.MultiheadAttention, which specialises the sequence
length under both the TorchScript and dynamo exporters; the resulting graph fails
outright at any other length. These graphs were exported after swapping in a shape-dynamic
attention that reuses the same weights and reproduces the original bit-identically
(max abs err 0.0). Verified length-independent from T=50 to T=1500, and end-to-end
against the upstream pipeline at correlation 0.99998752.
Usage
from audiosronnx import load_denoise
dn = load_denoise("mpsenet") # dns checkpoint
dn = load_denoise("mpsenet", model="vb") # VoiceBank+DEMAND checkpoint
clean, rate = dn.denoise("noisy.wav") # -> (float32 mono, 16000)
License
MIT, inherited from the upstream MP-SENet release.