Neural backends

The core of voicequal is pure numpy/scipy and runs in about 10 ms per clip. The neural extra adds reference-grade neural quality predictors behind the same interface, so you can compare against them, calibrate to them, or escalate to them when the cheap path is not enough.

pip install 'voicequal[neural]'

DNSMOS

DNSMOS (Reddy, Gopal, Cutler; Microsoft, ICASSP 2021/2022) is a non-intrusive predictor of the three ITU-T P.835 mean-opinion scores. It needs no clean reference.

Score	Meaning	Range
sig	Speech signal quality	1..5
bak	Background intrusiveness (higher = less intrusive)	1..5
ovrl	Overall quality	1..5

from voicequal.neural import DNSMOS

scorer = DNSMOS()  # downloads the 1.1 MB ONNX model once
scores = scorer.score_file("clip.wav")  # or scorer(samples, sample_rate)
print(scores.sig, scores.bak, scores.ovrl)

The model file (sig_bak_ovr.onnx, MIT licence) is fetched from the DNS-Challenge repository on first use, verified against a pinned SHA-256, and cached in ~/.voicequal/models/. Set VOICEQUAL_MODEL_DIR to move the cache, or pass model_path= to use a file you already have.

Inference mirrors Microsoft's reference script: clips shorter than 9.01 s are tiled, longer clips are scored in 9.01 s windows hopping 1 s, and raw outputs go through the published polynomial mapping before averaging.

DNSMOS is a speech model

It does not transfer to singing. On voicequal's VocalSet benchmark it scores clean sung vowels around 1.1 on every axis, the same as white noise. That is why the DSP-vs-DNSMOS comparison in the benchmark uses a public speech set.

When to use which

Path	Latency (3.6 s clip)	Needs	Best for
DSP (core)	~10 ms	numpy, scipy	Live mic checks, gating before ASR, singing
DNSMOS	~500 ms	onnxruntime, 1 MB model	Offline scoring of speech, human-MOS proxy

dnsmos_available() tells you at runtime whether the extra is installed.