Analysis and Mitigation of Vocal Effort Variations in Speaker Recognition

MAHESH KUMAR NANDWANA; DIEGO CASTÁN; LUCIANA FERRER; MITCHELL MCLAREN; AARON LAWSON

Brighton

Congreso; ICASSP 2019; 2019

IEEE

In this work, we assess the impact of vocal effort on discrimina-tion and calibration performance of a state-of-the-art speaker recog-nition system. We analyze three levels of vocal effort (low, nor-mal, and high) from the SRI-FRTIV corpus. We use a deep neuralnetwork (DNN) speaker embeddings system with probabilistic lin-ear discriminant analysis (PLDA) and find that vocal effort variationsignificantly degrades system performance. We apply both mixturePLDA (mix-PLDA) and trial-based calibration with condition PLDAsimilarity (TBC-CPLDA) to improve system robustness. Our pro-posed approaches resulted in 18% and 33% relative improvement indiscrimination and calibration performance respectively on the SRI-FRTIV corpus.