This paper describes the application of deep neural networks (DNNs), trained to discriminate among speakers, to improving performance in text-independent speaker verification. Activations from the bottleneck layer of these DNNs are used as features in an i-vector based speaker verification system. The features derived from this network are thought to be more robust with respect to phonetic variability, which is generally considered to have a negative impact on speaker verification performance. The verification performance using these features is evaluated on the 2012 NIST SRE core-core condition with models trained from a subset of the Fisher and Switchboard conversational speech corpora. It is found that improved performance, as measured by the minimum detection cost function (minDCF), can be obtained by appending speaker discriminative features to the more widely used mel-frequency cepstrum coefficients.

The use of a graph embedding framework is investigated as a regularization technique in the expectation-maximization (EM) algorithm applied to automatic speech recognition (ASR). The technique is motivated by the fact that graph em-beddings of feature vectors have been shown to provide useful characterizations of the underlying manifolds on which these features lie. Incorporating intrinsic graphs that describe these manifolds in the optimization criteria for the EM algorithm has the effect of constraining the solution space in a way that preserves the local structure of the data. Graph embedding based regularization is applied here to estimating parameters in constrained maximum likelihood linear regression (CMLLR) speaker adaptation in continuous density hidden Markov model (CDHMM) based ASR. CMLLR adaptation has been widely used as a maximum likelihood procedure for reducing mismatch between a given HMM model and utterances from an unknown speaker through a linear feature space transformation. However, there is no guarantee that CMLLR transformations will preserve the relationships of the feature vectors along this manifold. It is argued here that graph embedding based regularization will preserve this structure. The impact of this approach on ASR performance is evaluated for unsupervised speaker adaptation on two large vocabulary speech corpora.

A two-stage speaker adaptation approach is proposed for the subspace Gaussian mixture model (SGMM) [1] in large vocabulary automatic speech recognition (ASR). The SGMM differs from the more well known continuous density hidden Markov model (CDHMM) in that a large portion of the SGMM parameters are dedicated to shared full covariance Gaussian subspace parameters and a relatively small number of parameters are used for state dependent projection vectors. Both model space and feature space adaptation are investigated. First, an efficient regression based approach for subspace vector adaptation (SVA) is presented. Second, an efficient approach is presented for feature space adaptation using constrained maximum likelihood linear regression (CMLLR) in the SGMM. While both of these adaptation scenarios have previously been investigated in the context of the SGMM [2, 3], a more efficient and numerically stable procedure is presented here for estimating the parameters of the regression based transformations. Both transformation matrices are obtained using an optimization technique that iteratively updates the rows of the regression matrices. It is shown that using these feature space and model space approaches for unsupervised speaker adaptation provides complementary improvements in SGMM based ASR word accuracy.

This paper presents a study of two acoustic speaker adaptation techniques applied in the context of the subspace Gaussian mixture model (SGMM) for automatic speech recognition (ASR). First, a model space linear regression based approach is presented for adaptation of SGMM state projection vectors and is referred to as subspace vector adaptation (SVA). Second, an easy to implement realization of constrained maximum likelihood linear regression (CMLLR) is presented for feature space adaptation in the SGMM. Numerically stable procedures for row-by-row estimation of the regression based transformation matrices are presented for both SVA and CMLLR adaptation. These approaches are applied to SGMM models that are estimated using speaker adaptive training (SAT), a technique for estimating more compact speaker independent acoustic models. Unsupervised speaker adaptation performance is evaluated on conversational and read speech task domains and compared to unsupervised adaptation performance obtained using the hidden Markov model-Gaussian mixture model (HMM-GMM) in ASR. It is shown that the feature space and model space adaptation approaches applied to the SGMM provide complementary reductions in word error rate (WER) and provide lower WERs than that obtained using CMLLR adaptation for the HMM-GMM.

An approach is proposed for adapting subspace projection vectors in the subspace Gaussian mixturemodel (SGMM) [1]. Subword models in the SGMM are composed of states, each of which are parametrized using a small number of subspace projection vectors. It is shown here that these projection vectors provide a compact and well-behaved characterization of phonetic information in speech. A regression based subspace vector adaptation approach is proposed for adapting these parameters. The performance of this approach is evaluated for unsupervised speaker adaptation on two large vocabulary speech corpora.

In this paper the use of acoustic similarity of speech intervals for generating improved confidence scores for spoken term detection (STD) is investigated. A procedure based on acoustic dotplots which requires no training data is deployed for discovering similar speech intervals. A graph based random walk algorithm incorporates acoustic similarity of hypothesized term occurrences for improving the corresponding confidence scores. The proposed approach is evaluated in an open vocabulary STD task defined on a lecture domain corpus. It is shown that updating the confidence scores in this fashion results in a significant increase in term detection performance of out of vocabulary search terms. A relative improvement of 12.9% in figure of merit was gained relative to that obtained from a baseline lattice based STD system.

The subspace Gaussian mixture model (SGMM) has been recently proposed as an acoustic modeling technique suitable for configuring multilingual speech recognition systems. It is attractive for this purpose since its parametrization allows its “shared” model parameters to be trained with data from multiple languages [1]. In this work, we report on the results of an experimental study carried out with the goal of improving native Spanish language speech recognition performance using an existing telephone speech corpus of English spoken by speakers of Spanish origin. Compensation for sources of acoustic variability between Spanish and English language data sets was found to be important in obtaining good multilingual ASR performance. We conclude with a discussion about the notion of acoustic similarity between the state dependent parameters of the SGMM, and its possible use in effectively modelling pronunciation variation.