Speaker verification calibration

SPARC

Frozen encoders. Calibrated scores.

A small residual on top of a frozen context-affine anchor turns speaker-verification scores into calibrated log-likelihood ratios, using duration, voice-activity and language context.

Get started View on GitHub

LLR = anchor + Δ. Because Δ starts at zero, training begins exactly at the anchor.

Better calibrated.
Across the board.

Tested on four corpora, six frozen encoders and their fusion — 28 conditions in all.

Panels improved

8 of 8

Lower C_llr than the context-affine baseline in every panel, and in 26 of 28 conditions.¹

Largest gain

9.0%

Lower C_llr than context affine on NISP fusion. TidyVoice fusion: 7.0%.¹

Against Variance-Gamma

18.7%

Lower C_llr on the TidyVoice single-encoder panel, with multiplicity-adjusted support.²

Subgroups

43 of 44

Improved when trials are split by duration, mismatch, language match and anchor confidence.³

Noisy context

32 of 32

Simulated VAD and language-ID error settings that keep the fusion gain.³

Speed

1.6×

Faster than Variance-Gamma on the same CPU, with one thread.⁴

Test C_llr · lower is better	Context affineBaseline	SPARC	Change
TidyVoice single	0.2518	0.2468	−2.0%
TidyVoice fusion	0.0926	0.0861	−7.0%
NISP single	0.1752	0.1679	−4.2%
NISP fusion	0.0371	0.0337	−9.0%
LibriSpeech single	0.1069	0.1062	−0.6%
LibriSpeech fusion	0.0368	0.0350	−4.9%
VoxCeleb2 single	0.2144	0.2112	−1.5%
VoxCeleb2 fusion	0.0644	0.0627	−2.6%

Single is the mean of six single-encoder conditions. SPARC values average the selected family's seeds.¹

Anchor first.
Then correct.

SPARC learns only what a strong affine reference misses — and keeps the reference when a correction does not help.

1 · Fit

A frozen anchor.

A source-affine map and a context-affine anchor are fitted on labeled calibration data, then frozen. Every residual starts at zero.

2 · Select

Earned capacity.

Validation chooses among 17 families — the anchor itself, 8 polynomial and 8 small neural residuals — only if C_llr and EER do not get worse.

3 · Deploy

One small file.

A JSON bundle of about 15 KB with at most 289 residual parameters. Prediction checks encoder order and context schema.

Up and running
in a minute.

The synthetic example fits the full procedure, saves a bundle and reloads it. No audio, pretrained weights or datasets needed.

Terminal

pip install -r requirements.txt
python -m examples.synthetic_demo --output-dir outputs/synthetic

Python

from delivery import Plugin, Q_COLUMNS

names = ("encoder_a", "encoder_b")
Plugin.fit(calibration, validation, names, output_path="sparc.json")
llr = Plugin.load("sparc.json").predict(scores, context, names, context_names=Q_COLUMNS)

Input format Method details