Sourjya Sarkar

Stochastic feature compensation methods for speaker verification in noisy environments

Abstract This paper explores the significance of stereo-based stochastic feature compensation (SFC) methods for robust speaker verification (SV) in mismatched training and test environments. Gaussian Mixture Model (GMM)-based SFC methods developed in past has been solely restricted for speech recognition tasks. Application of these algorithms in a SV framework for background noise compensation is proposed in this paper. A priori knowledge about the test environment and availability of stereo training data is assumed. During the training phase, Mel frequency cepstral coefficient (MFCC) features extracted from a speakers noisy and clean speech utterance (stereo data) are used to build front end GMMs. During the evaluation phase, noisy test utterances are transformed on the basis of a minimum mean squared error (MMSE) or maximum likelihood (MLE) estimate, using the target speaker GMMs. Experiments conducted on the NIST-2003-SRE database with clean speech utterances artificially degraded with different types of additive noises reveal that the proposed SV systems strictly outperform baseline SV systems in mismatched conditions across all noisy background environments.

Speaker verification in noisy environment using GMM supervectors

This paper explores the GMM-SVM combined approach for Text-Independent speaker verification in noisy environment. In recent years supervectors constructed by stacking the means of adapted Gaussian Mixture Models (GMMs) have been used successfully for deriving sequence kernels. Support Vector Machines (SVMs) trained using such kernels provide further improvement in classification accuracy. Analysis of the behavior of such hybrid systems towards simulated noisy data is the object of our study. In our work we have used the KL-divergence and GMM-UBM mean interval kernels for SVM training. All experiments are conducted on NIST-SRE-2003 database with training and test utterances degraded by noises (car, factory & pink) collected from the NOISEX-92 database, at 5dB & 10dB SNRs. A significant improvement of performance is observed in comparison to the traditional GMM-UBM based system.

Amplitude estimation of a sinusoid buried in heavy noise using stochastic resonance

This paper investigates four methods to estimate the amplitude of a sinusoid buried in noise, using the phenomenon of stochastic resonance (SR) to obtain the best estimate. SR is a cooperative effect in which the power in the broadband part of the spectrum is fed in to the output power of the signal frequency. We use a 3-level quantizer in our experiments to quantize the sinusoid buried in noise after variance-controlled AWGN is added to it We estimate the amplitude at the optimum variance, which is obtained from the experimentally obtained SNR data.

Supervector-based approaches in a discriminative framework for speaker verification in noisy environments

This paper explores the robustness of supervector-based speaker modeling approaches for speaker verification (SV) in noisy environments. In this paper speaker modeling is carried out in two different frameworks: (i) Gaussian mixture model-support vector machine (GMM-SVM) combined method and (ii) total variability modeling method. In the GMM-SVM combined method, supervectors obtained by concatenating the mean of an adapted speaker GMMs are used to train speaker-specific SVMs during the training/enrollment phase of SV. During the evaluation/testing phase, noisy test utterances transformed into supervectors are subjected to SVM-based pattern matching and classification. In the total variability modeling method, large size supervectors are reduced to a low dimensional channel robust vector (i-vector) prior to SVM training and subsequent evaluation. Special emphasis has been laid on the significance of a utterance partitioning technique for mitigating data-imbalance and utterance duration mismatches. An adaptive boosting algorithm is proposed in the total variability modeling framework for enhancing the accuracy of SVM classifiers. Experiments performed on the NIST-SRE-2003 database with training and test utterances corrupted with additive noises indicate that the aforementioned modeling methods outperform the standard GMM-universal background model (GMM-UBM) framework for SV. It is observed that the use of utterance partitioning and adaptive boosting in the speaker modeling frameworks result in substantial performance improvements under degraded conditions.

Robust Speaker Verification: A Review

This chapter provides an overview of various feature and model-based approaches developed in past for robust speaker recognition. The advantages and disadvantages of some standard methods applied for robust speaker verification tasks have been highlighted. The main focus is to summarily introduce popular state-of-the-art techniques adopted for enhancing speaker verification performance in noisy conditions.

Significance of utterance partitioning in GMM-SVM based speaker verification in varying background environment

This paper explores the GMM-SVM combined approach for text-independent speaker verification in rapidly varying environmental noise. For mitigating the effect of mismatched training and test utterance length and countering the impact of data-imbalance in SVM scoring, we partition each full-length enrollment utterance into a number of sub-utterances and derive a GMM supervector from each of them prior to SVM training. Experiments conducted on the NIST-SRE-2003 database demonstrate that the GMM-SVM system with partitioned utterances outperform the conventional GMM-SVM based speaker verification system. The noisy background is simulated by degrading non-overlapping segments of each training and test utterance of the NIST-SRE-2003 by additive noises (car, factory, pink & white) collected from the NOISEX-92 database, at OdB, 5dB, 7dB & 10dB SNRs respectively. An average performance improvement of 6.41% and 10.56% EER across all SNRs is observed in comparison to the traditional GMM-SVM and GMM-UBM based systems, respectively.

Multilingual speaker recognition on Indian languages

In this paper we explore the performance of multilingual speaker recognition systems developed on the IITKGP-MLILSC speech corpus. Closed-set speaker identification and speaker verification experiments are individually conducted on 13 widely spoken Indian languages. In particular, we focus on the effect of language mismatch in the speaker recognition performance of individual languages and all languages together. The standard GMM-based speaker recognition framework is used. While the average language-independent speaker identification rate is as high as 95.21%, an average equal error rate of 11.71% shows scope for further improvement in speaker verification performance.

Stochastic Feature Compensation for Robust Speaker Verification

This chapter explores the impact of standard stereo-based stochastic feature compensation (SFC) methods for robust speaker verification in uniform noisy environments. In this work, SFC using independent as well as joint probability models are explored for compensating the effect of noise. Integration of a SFC stage in the GMM-UBM framework is proposed for speaker verification evaluation under mismatched conditions.

Speaker Verification in Noisy Environments Using Gaussian Mixture Models

This chapter explores the behavior of Gaussian Mixture Models (GMMs) for speaker verification in noisy environments. Specifically, the performance of an acoustic modeling framework (namely GMM-UBM) using speaker-dependent GMMs and a speaker-independent Universal Background Model (UBM), is studied for simulated noisy backgrounds. Significance of a feature mapping technique using multiple UBMs for compensating background noise is explored. The speaker verification systems explored in this chapter serve the purpose of baselines considered for comparison and analyzing the performance improvements of the proposed methods in the remaining chapters.

Robust Speaker Modeling for Speaker Verification in Noisy Environments

The present chapter explores robust speaker modeling methods for speaker verification in noisy environment. The focus is specifically laid on building hybrid classifiers based on the combination of generative and discriminative models (e.g., Gaussian Mixture Models (GMMs) and Support Vector Machines (SVMs)). For improving the performance of the proposed speaker verification systems, utterance partitioning methods are used. The discussion is closely followed by state-of-the-art variants of GMM supervector based approaches (i.e., i-vectors) and algorithms for combining robust classifiers.