VTLN-warped Gaussian posteriorgram for QbE-STD

Abstract

Abstract-Vocal Tract Length Normalization (VTLN) is a very important speaker normalization technique for speech recognition tasks. In this paper, we propose the use of Gaussian posteriorgram of VTLN-warped spectral features for a Queryby-Example Spoken Term Detection (QbE-STD). This paper presents the use of a Gaussian Mixture Model (GMM) framework for estimation of VTLN warping factor. This GMM framework does not require phoneme-level transcription and hence, it can be useful for unsupervised tasks. We propose the iterative approach for VTLN warping factor estimation with two GMM training approaches, namely, Expectation-Maximization (EM) and Deterministic Annealing-Expectation Maximization (DAEM). The VTLN-warped Gaussian posteriorgram gave the better QbE-STD performance. The performance of TIMIT QbE-STD was investigated with different evaluation factors, such as a number of Gaussian components in GMM, various local constraints, and a number of iterations in VTLN warping factor estimation. VTLNwarped Gaussian posteriorgram reduces the speaker-specific variation in Gaussian posteriorgram and hence, it is expected to give better performance than Gaussian posteriorgram.

Unsupervised Vocal Tract Length Warped Posterior Features for Non-Parallel Voice Conversion

“…Thus, both the set have non-overlapping utterances. 25-D Mel Cepstral Coefficients (MCCs) (including the 0 th coefficient) and 39-D PLP features (including ∆ and ∆∆) as suggested in [28], [33]. 1-D F0 for each frame (with 25 ms frame duration, and 5 ms frame shift) have been extracted.…”

“…The maximum likelihood search is performed to obtain suitable speaker-specific VTLN warping factor [31,32]. The procedure of VTL-warped Gaussian posteriorgram is as follows [28], [33]:…”

“…speaker-independent GMM can be captured by VTLN warping factor. In order to estimate the VTLN warping factor, we follow the approach as suggested in our previous work [28], [33]. VTLN warping factor is estimated from the sets of different warped feature vector sequences X α , with different values of α.…”

“…Figure 1: Schematic block diagram of extracting VTL-warped posterior features. After [28], [33]. Figure 1 shows the block diagram of warped Gaussian posteriorgram computation using VTLN warping factor and speakerindependent GMM.…”

“…This new speaker-independent GMM is further used to estimate VTLN warping factor and warped acoustic features are again used for re-training. This procedure is iteratively executed for five iterations (as suggested in [28], [33]). Speakerpairs data from the VCC 2016 database is used to train speakerindependent GMM for the VTLN posteriorgram.…”