Probabilistic feature-based transformation for speaker verification over telephone networks

Abstract

Feature transformation aims to reduce the effects of channel-and handset-distortion in telephone-based speaker verification. This paper compares several feature transformation techniques and evaluates their verification performance and computation time under the 2000 NIST speaker recognition evaluation protocol. Techniques compared include feature mapping (FM), stochastic feature transformation (SFT), blind stochastic feature transformation (BSFT), feature warping (FW), and short-time Gaussianization (STG). The paper proposes a probabilistic feature mapping (PFM) in which the mapped features depend not only on the top-1 decoded Gaussian but also on the posterior probabilities of other Gaussians in the root model. The paper also proposes speeding up the computation of PFM and BSFT parameters by considering the top few Gaussians only. Results show that PFM performs slightly better than FM and that the fast approach can reduce computation time substantially. Among the approaches investigated, the fast BSFT (fBSFT) strikes a good balance between computational complexity and error rates, and FW and STG are the best in terms of error rates but with higher computational complexity. It was also found that fusion of the scores derived from systems using fBSFT and STG can reduce the error rate further. This study advocates that fBSFT, FW, and STG have the highest potential for robust speaker verification over telephone networks because they achieve good performance without any a priori knowledge of the communication channel.

A new adaptation approach to high-level speaker-model creation in speaker verification

“…For the acoustic GMM-UBM system [1], we applied several channel compensation techniques, including feature warping [40], Z-norm [41], short-time Gaussianization (STG) [42] and fast blind stochastic feature transformation (fBSFT) [43]. Acoustic scores S GMM-UBM were computed based on the loglikelihood ratio:…”

“…Speaker Detection Performance For (a), short-time Gaussianization (STG) and fast blind stochastic feature transformation (fBSFT) [43] were applied to the low-level features, and for (b) feature warping was applied.…”

Speaker Verification via High-Level Feature Based Phonetic-Class Pronunciation Modeling

“…The MFCCs and delta MFCCs were concatenated to form 38-dimensional feature vectors. Cepstral mean subtraction (CMS), fast blind stochastic features transformation (fBSFT) [26], [3] and short-time Gaussianization (STG) [27] were applied to the MFCCs to remove channel effects.…”

“…Making the low-level features robust, however, does not come without a price. It has been shown recently that using STG and fast BSFT as feature preprocessors requires 52 seconds to process a 53-second utterance on a Pentium IV 3.2GHz CPU, whereas processing the same utterance by the less powerful cepstral mean subtraction takes only 0.02 seconds [3].…”

“…However, the lack of robustness to mismatched conditions remains a serious problem. Although approaches such as feature transformation [2], [3], model transformation [4], and score normalization [5] have shown promise in reducing the mismatches, these methods have almost reached their limit in terms of error rate reduction.…”