Robustness of Quality-based Score Calibration of Speaker Recognition - PowerPoint PPT Presentation

Robustness of Quality-based Score Calibration of Speaker Recognition Systems with respect to low-SNR and short-duration conditions Andreas Nautsch ⋆ , Rahim Saeidi † , Christian Rathgeb ⋆ , Christoph Busch ⋆ ⋆ Hochschule Darmstadt, CRISP, CASED, da/sec Security Research Group † Department of Signal Processing and Acoustics, Aalto University, Finland Odyssey’16, Bilbao, 24.06.2016 Nautsch, Saeidi, Rathgeb, Busch Quality-based Score Calibration / Odyssey’16, Bilbao, 24.06.2016 1/3

Motivation Quality-based Calibration and low-Quality Robustness Calibration schemes in unconstrained conditions ◮ Conventional: train on full/clean ⇒ use on all df: 2 ◮ Matched: train and use for each condition df: 2 C ◮ Quality Measure Function: duration/SNR/. . . df: 3 in linear fusion: S ′ = w 0 + w 1 S + w 2 Q ( λ r ef , λ p rb ) Unified Audio Characteristics ◮ Gaussian models for conditions on raw i-vector domain as vector representation of quality estimates � q df: C 2 + 3 ◮ Bi-linear combination matrix: w 2 � q r ef Σ � q p rb ◮ Proposed: Function of Quality Estimates (FQEs) df: 3 e.g., by cosine distance: S ′ = w 0 + w 1 S + w 2 cos ( � q r ef , � q p rb ) Nautsch, Saeidi, Rathgeb, Busch Quality-based Score Calibration / Odyssey’16, Bilbao, 24.06.2016 2/3

Results Robustness Analyses in 55 combined conditions Excerpt: 100 C mc × 10 3 50 0 B Conventional B d B Q qvec d B 5 d 0 B / d Q 1+2 / 5 B l d l n l / 0 u 5 d l a B u s / f B Q 2 0 / 0 e f s d B 4 0 s / l d c 0 0 n 4 s d Q 1 / B 2 0 2 5 a s 0 B 2 / 1 e d 0 B 1 l s / d Matched 1 c 0 0 / d s / 2 5 low SNR 1 0 s s 0 / 1 0 1 5 1 s / 1 5 s / 5 s 5 short duration Calibration Condition Nautsch, Saeidi, Rathgeb, Busch Quality-based Score Calibration / Odyssey’16, Bilbao, 24.06.2016 3/3