Intra-class covariance adaptation in PLDA back-ends for speaker verification

Abstract

Multi-session training conditions are becoming increasingly common in recent benchmark datasets for both textindependent and text-dependent speaker verification. In the state-of-the-art i-vector framework for speaker verification, such conditions are addressed by simple techniques such as averaging the individual i-vectors, averaging scores, or modifying the Probabilistic Linear Discriminant Analysis (PLDA) scoring hypothesis for multi-session enrollment. The aforementioned techniques fail to exploit the speaker variabilities observed in the enrollment data for target speakers. In this paper, we propose to exploit the multi-session training data by estimating a speaker-dependent covariance matrix and updating the intra-speaker covariance during PLDA scoring for each target speaker. The proposed method is further extended by combining covariance adaptation and score averaging. In this method, the individual examples of the target speaker are compared against the test data as opposed to an averaged ivector, and the scores obtained are then averaged. The proposed methods are evaluated on the NIST SRE 2012 dataset. Relative improvements of up to 29% in equal error rate are obtained.

Voice liveness detection using phoneme-based pop-noise detector for speaker verification

“…Automatic speaker verification (ASV) is an easy-to-use biometric authentication using speech. State-of-the-art ASV systems, as i-vector-based [2][3][4] and probabilistic linear discriminant analysis (PLDA)-based ones [5][6][7], have achieved reliable performances and are expected to be in practical use. Meanwhile, performances of speech synthesis algorithms such as text-to-speech (TTS) systems [8][9][10] and voice conversion systems [11,12] have been significantly improved.…”

Robust and Discriminative Speaker Embedding via Intra-Class Distance Variance Regularization

“…To extract i-vectors, Baum-Welch statistics are computed from a Gaussian Mixture Model-Universal Background Model (GMM-UBM), which is learned using a sequence of feature vectors. I-vectors then can be used to compare utterances directly using cosine similarity or probabilistic linear discriminant (PLDA) [14,15,16]. To improve upon i-vectors, deep neural networks (DNNs) have been first applied to gradually replace each step in computing i-vectors traditional speaker recognition systems [17,18].…”