Distilled quality model

voicequal ships a small learned model that predicts the three ITU-T P.835 scores, SIG (speech quality), BAK (background intrusiveness) and OVRL (overall), from a clip. It is a student of DNSMOS, Microsoft's neural P.835 predictor: it was trained to reproduce DNSMOS's outputs, so it gives you DNSMOS-like scores with no onnxruntime, no model download, and sub-millisecond inference after feature extraction.

from voicequal.models import predict_mos

mos = predict_mos(samples, sample_rate=16000)
print(mos.sig, mos.bak, mos.ovrl)  # 1..5 each

How it is built

Step	What
Features	169 values per clip: voicequal's 9 DSP aggregates (SNR estimate, HNR, energy SNR, room level, spectral SNR, flatness, concentration, temporal variance, clipping) plus mean / std / 10th / 90th percentile of a 40-band log-mel spectrogram. Pure numpy. See voicequal.features.
Teacher	DNSMOS sig_bak_ovr.onnx run on every training clip.
Data	VoiceBank-DEMAND train shard 0 (28 speakers): noisy clips plus a sample of their clean originals. Evaluated on the VoiceBank-DEMAND test split (2 speakers never seen in training).
Student	Standardise → Linear(169→64) → ReLU → Linear(64→32) → ReLU → Linear(32→3), clipped to 1..5. About 13k parameters.
Artefacts	src/voicequal/models/quality_mlp.npz (numpy weights, shipped in the wheel) and benchmarks/distill/quality_mlp.onnx (same network for browsers and other runtimes).

Reproduce with:

pip install 'voicequal[neural]' pyarrow scikit-learn onnx
python benchmarks/distill/label.py    # DNSMOS labels + features -> dataset.npz
python benchmarks/distill/train.py    # fits, evaluates, exports npz + onnx

How close to the teacher

Evaluated on the VoiceBank-DEMAND test split: 824 noisy clips plus their 824 clean originals, two speakers the model never saw. "Baseline" is a straight line fitted from snr_estimate alone, to show what the mel features add.

Target	Student vs DNSMOS, Pearson	Spearman	MAE (MOS)	Baseline Pearson	Baseline MAE
BAK (background)	0.93	0.85	0.19	0.76	0.45
OVRL (overall)	0.87	0.79	0.18	0.68	0.35
SIG (speech)	0.65	0.44	0.19	0.36	0.38

Against the true mixing SNR of the noisy test clips, the student's OVRL ranks at +0.56 and the teacher's at +0.57: the student reproduces the teacher's view, including where the teacher disagrees with raw SNR.

	Student (numpy)	DNSMOS (onnxruntime)
Parameters	13,059	~1.1 M
Size	51 KB	1.1 MB
Inference per clip, after features	~70 µs	~180 ms uncontended, ~880 ms on a busy machine
Feature extraction	~25 ms	included above

SIG is the weak axis. DNSMOS's own SIG is the noisiest of its three outputs on this data, and a clip-level feature vector loses the fine-grained speech distortion cues it keys on. BAK and OVRL are what most pipelines gate on, and those are close to the teacher.

Training regularisation was chosen by a small sweep on the test split (L2 penalty 3.0, hidden 64/32). That is a mild form of test-set tuning; a held-out validation split from the remaining train shards is the right next step.

What it is not

• Not a human-rated model. Its targets are DNSMOS predictions, not listener scores. It can be no better than DNSMOS, and it adds its own error on top.
• Not for singing. DNSMOS scores clean singing like white noise, so the student learned nothing useful about singing. For singing, use the tier and snr_estimate from assess().
• Not a replacement for the tier. The tier decision stays the rule-based ladder, which is explainable and works on both speech and singing. The MOS estimate is an extra output for pipelines that speak P.835.