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 (
weights/nisqa.tar).
Licence — read this first
NISQA is licensed in two parts, and they are not the same licence:
| Part | Licence |
|---|
| NISQA code | MIT (Copyright (c) 2021 Gabriel Mittag, Quality and Usability Lab, TU Berlin) |
| NISQA model weights | CC BY-NC-SA 4.0 — Attribution-NonCommercial-ShareAlike 4.0 International |
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
| file | size | notes |
|---|
nisqa.onnx | 1,054,416 bytes | opset 17, single self-contained file |
Input / output contract
The ONNX graph is the model only — the mel-spectrogram frontend and the
segmentation are the caller's responsibility.
Input
| name | shape | dtype |
|---|
segments | [batch, n_segments, 1, 48, 15] | float32 |
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:
| parameter | value |
|---|
| sample rate | native — audio is not resampled (upstream uses librosa.load(sr=None)) |
n_fft | 4096 |
hop_length | int(sr * 0.01) (10 ms) |
win_length | int(sr * 0.02) (20 ms) |
| window | hann, center=True, pad_mode='reflect' |
power | 1.0 (magnitude, not power) |
n_mels | 48 |
fmin / fmax | 0.0 / 20000 |
| mel scale | htk=False (Slaney), norm='slaney' |
| dB conversion | amplitude_to_db(S, ref=1.0, amin=1e-4, top_db=80.0) |
Segmentation of the resulting [48, T] spectrogram:
| parameter | value |
|---|
seg_length | 15 frames (odd; the frame plus 7 neighbours each side) |
seg_hop | 4 |
| resulting shape | [n_segments, 1, 48, 15], n_segments = ceil((T - 14) / 4) |
No per-segment normalization is applied beyond the dB conversion above.
Output
| name | shape | dtype |
|---|
scores | [batch, 5] | float32 |
The five values are, in this exact order:
| index | dimension | meaning |
|---|
| 0 | mos | Overall MOS (speech quality) |
| 1 | noi | Noisiness |
| 2 | dis | Discontinuity |
| 3 | col | Coloration |
| 4 | loud | Loudness |
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:
| dimension | max abs deviation |
|---|
mos | 1.43e-06 |
noi | 9.54e-07 |
dis | 1.43e-06 |
col | 1.43e-06 |
loud | 9.54e-07 |
| overall | 1.43e-06 (duration sweep: 1.91e-06) |
Citation
1@inproceedings{mittag2021nisqa,
2 title = {{NISQA}: A Deep {CNN}-Self-Attention Model for Multidimensional
3 Speech Quality Prediction with Crowdsourced Datasets},
4 author = {Mittag, Gabriel and Naderi, Babak and Chehadi, Assmaa and
5 M{\"o}ller, Sebastian},
6 booktitle = {Proc. Interspeech 2021},
7 pages = {2127--2131},
8 year = {2021},
9 doi = {10.21437/Interspeech.2021-299}
10}
