Combining stochastic feature transformation and handset identification for telephone-based speaker verification

Abstract

The performance of telephone-based speaker verification systems can be severely degraded by the acoustic mismatch caused by telephone handsets. This paper proposes to combine a handset selector with stochastic feature transformation to reduce the mismatch. Specifically, a GMM-based handset selector is trained to identify the most likely handset used by the claimants, and then handset-specific stochastic feature transformations are applied to the distorted feature vectors. To overcome the non-linear distortion introduced by telephone handsets, a 2nd-order stochastic feature transformation is proposed. Estimation algorithms based on the stochastic matching technique and the EM algorithm are derived. Experimental results based on 150 speakers of the HTIMIT corpus show that the handset selector is able to identify the handsets accurately (98.3%), and that both linear and non-linear transformation reduce the error rate significantly (from 12.37% to 5.49%).

Probabilistic Fusion of Sorted Score Sequences for Robust Speaker Verification

“…We used a GSM speech codec to transcode the HTIMIT corpus [9] and applied the resulting transcoded speech in a speaker verification experiment similar to [10] and [11]. HTIMIT was obtained by playing a subset of the TIMIT corpus through 9 different telephone handsets and one Sennheizer head-mounted microphone.…”

“…As a result, there were handset-and coder-mismatches between speaker models and verification utterances. We used stochastic feature transformation with handset identification [10][13] to compensate the mismatches. We assumed that a claimant will be asked to utter two sentences during a verification session.…”

Cluster-Dependent Feature Transformation for Telephone-Based Speaker Verification

“…In our previous work [1], a handset selector is designed to identify the most likely handset used by the claimants. The handset's identity was then used to select the transformation parameters to recover the distorted speech.…”

“…with mixing coefficients ω X j , mean vectors µ X j and covariance matrices Σ X j derived from the clean speech of several speakers (ten speakers in this work), the maximum-likelihood estimates of ν can be iteratively computed via the expectation-maximization (EM) algorithm [4] as follows [1] …”

“…As a telephone number can associate with several handsets, determining the handset type based on a given phone number alone is not very reliable. Without a reliable method to identify the handset type, the approaches proposed in [1] and [3] become less useful.…”