ZipEnhancer-ONNX / README.md
PatiPix's picture
Publication de zipenhancer.onnx par PatiPix Studio
675ed98 verified
|
Raw History Blame Contribute Delete
2.06 kB
---
license: apache-2.0
tags:
- onnx
- speech-enhancement
- noise-suppression
- zipenhancer
---
# ZipEnhancer (16 kHz) — portable ONNX
Portable ONNX export of ZipEnhancer
([iic/speech_zipenhancer_ans_multiloss_16k_base](https://modelscope.cn/models/iic/speech_zipenhancer_ans_multiloss_16k_base)
on ModelScope, revision `bed374bb53e4a5ffc62093676bab408a56b0db48`, file `pytorch_model.bin`), the noise suppressor
that VoxCPM2 uses for its `denoise` option.
PatiPix Studio uses it for the optional « Remove background noise » step of its voice cloning.
Why a new export: the ONNX file published with the original model was optimized by ONNX Runtime for the processor
that prepared it (`com.microsoft.nchwc` operators), so it is not safe on other PCs. This export only uses standard
ONNX operators.
## Inputs and outputs
| Name | Shape | Meaning |
|---|---|---|
| `noisy_mag` (input) | [1, 201, T] | STFT magnitude, compressed (power 0.3) |
| `noisy_pha` (input) | [1, 201, T] | STFT phase |
| `amp_g`, `pha_g` (outputs) | [1, 201, T] | cleaned magnitude (compressed) and phase |
STFT as in ModelScope's `mag_pha_stft` / `mag_pha_istft`: 16 kHz audio normalised to unit RMS, 400 points, hop 100,
periodic Hann window, centered with reflect padding. Valid up to T = 961 frames (6 s) in one block, like the official
pipeline, which processes longer audio in 2 s chunks (0.25 s dropped on each side).
## How it was made
Official ModelScope 1.40.2 code, torch 2.10.0, TorchScript export (`dynamo=False`), opset 17, weights unchanged.
Checked on the two official noisy examples: amplitude difference between ONNX and PyTorch below 3e-6; the full chain
agrees with the official ModelScope pipeline at 58 to 61 dB (the limit of 16-bit audio).
## File
| File | Size (bytes) | SHA-256 |
|---|---|---|
| `zipenhancer.onnx` | 9,591,998 | `4ae0b92cd0d5933dc3f4896cd0ad62132e0249d4b6f4b33d95c7518a364d3c70` |
## License and credits
Apache License 2.0, like the original model (see `LICENSE`). Original work: ZipEnhancer by the Tongyi Lab (Alibaba).