TigreGotico/nisqa-onnx

NISQA (ONNX)

ONNX export of NISQAv2, the multi-dimensional no-reference speech quality model by Mittag et al. It predicts five perceptual quality dimensions from a single degraded speech signal, with no clean reference required.

This repository contains a derivative ONNX export of the model weights from [gabrielmittag/NISQA](https://github.com/gabrielmittag/NISQA) (weights/nisqa.tar).

Licence — read this first

NISQA is licensed in two parts, and they are not the same licence:

The ONNX file in this repository is a derivative of the weights, so it inherits CC BY-NC-SA 4.0:

• —NonCommercial — you may not use it for commercial purposes.
• —ShareAlike — derivatives must be distributed under the same licence.
• —Attribution — credit the original authors.

The upstream weights licence is included verbatim as `LICENSE_model_weights`. This repository is kept separate from our MIT-licensed metric models precisely so that the licence tag tells the truth.

Files

Input / output contract

The ONNX graph is the model only — the mel-spectrogram frontend and the segmentation are the caller's responsibility.

Input

n_segments is a dynamic axis; variable-length audio works without re-export. Unlike upstream, the graph takes no `n_wins` input and expects no padding — feed exactly the real segments and every mask is implicitly all-true.

Frontend (must be reproduced exactly by the caller)

Mel-spectrogram, computed with librosa semantics:

Segmentation of the resulting [48, T] spectrogram:

No per-segment normalization is applied beyond the dB conversion above.

Output

The five values are, in this exact order:

All five are on the MOS scale (roughly 1–5). No post-processing or polynomial correction is required.

Export notes

Two upstream constructs do not survive a naive trace and were replaced with mathematically identical, shape-agnostic equivalents:

• —F.adaptive_max_pool2d — not exportable when the input size is not statically visible. The three pools have architecturally fixed spatial sizes (48×15 → 24×7 → 12×5 → 6×3), so the adaptive windows are enumerated explicitly.
• —nn.MultiheadAttention — traces to Reshape nodes with the trace-time sequence length baked in as a constant, which breaks the dynamic segment axis. NISQAv2 uses nhead=1, so it is replaced with a plain scaled dot-product attention over the packed in_proj weights.

Both were verified numerically against the unmodified upstream padded forward pass (which uses pack_padded_sequence and n_wins masking).

Verified parity

15 real speech clips spanning the quality range, plus a 1 s / 3 s / 10 s duration sweep exercising the dynamic axis. Maximum absolute deviation from the PyTorch reference, per dimension:

Citation

@inproceedings{mittag2021nisqa,
  title     = {{NISQA}: A Deep {CNN}-Self-Attention Model for Multidimensional
               Speech Quality Prediction with Crowdsourced Datasets},
  author    = {Mittag, Gabriel and Naderi, Babak and Chehadi, Assmaa and
               M{\"o}ller, Sebastian},
  booktitle = {Proc. Interspeech 2021},
  pages     = {2127--2131},
  year      = {2021},
  doi       = {10.21437/Interspeech.2021-299}
}

Upstream: <https://github.com/gabrielmittag/NISQA>