Suryakanth V. Gangashetty

Online text-independent speaker verification system using autoassociative neural network models

In this paper, we present an approach based on autoassociative neural network (AANN) model for online implementation of text-independent speaker verification system. The distribution capturing ability of an AANN model is exploited to build the speaker model. A rank-based approach using individual background models is used to verify the claim of the speaker.

Extraction of fixed dimension patterns from varying duration segments of consonant-vowel utterances

Classification models based on multilayer perceptron (MLP) or support vector machine (SVM) have been commonly used for complex pattern classification tasks. These models are suitable for classification of fixed dimension patterns. However, durations of consonant-vowel (CV) utterances vary not only for different classes, but also for a particular CV class. It is necessary to develop a method for representing the CV utterances by patterns of fixed dimension. For CV utterances, vowel onset point (VOP) is the instant at which the consonant part ends and the vowel part begins. Important information necessary for classification of CV utterances is present in the region around the VOP. A segment of fixed duration around the VOP can be processed to extract a pattern of fixed dimension to represent a CV utterance. Accurate detection of vowel onset points is important for recognition of CV utterances of speech. In this paper, we propose an approach for detection of VOP, based on dynamic time alignment between a reference pattern of a CV class and the pattern of an utterance of that class. The results of studies show that the hypothesised VOPs using the proposed approach have less deviation from their actual locations.

Spotting multilingual consonant-vowel units of speech using neural network models

Multilingual speech recognition system is required for tasks that use several languages in one speech recognition application. In this paper, we propose an approach for multilingual speech recognition by spotting consonant-vowel (CV) units. The important features of spotting approach are that there is no need for automatic segmentation of speech and it is not necessary to use models for higher level units to recognise the CV units. The main issues in spotting multilingual CV units are the location of anchor points and labeling the regions around these anchor points using suitable classifiers. The vowel onset points (VOPs) have been used as anchor points. The distribution capturing ability of autoassociative neural network (AANN) models is explored for detection of VOPs in continuous speech. We explore classification models such as support vector machines (SVMs) which are capable of discriminating confusable classes of CV units and generalisation from limited amount of training data. The data for similar CV units across languages are shared to train the classifiers for recognition of CV units of speech in multiple languages. We study the spotting approach for recognition of a large number of CV units in the broadcast news corpus of three Indian languages.

A speech-based conversation system for accessing agriculture commodity prices in Indian languages

We demonstrate a speech based conversation system under development for information access by farmers in rural and semi-urban areas of India. The challenges are that the system should take care of the significant variations in the pronunciation and also the highly natural and hence unstructured dialog in the usage of the system. The focus of this study is to develop a conversational system which is adaptable to the users over a period of time, in the sense that fewer interactions with the system to get the required information. Some other novel features of the system include multiple decoding schemes and accountability of the wide variations in dialog, pronunciation and environment. A video demonstrating the Mandi information system is available at http://speech.iiit.ac.in/index.php/demos.html

Combining evidence from multiple classifiers for recognition of consonant-vowel units of speech in multiple languages

In this paper, we present studies on combining evidence from multiple classifiers to recognize a large number of consonant-vowel (CV) units of speech. Multiple classifier systems may lead to a better solution to the complex speech recognition tasks, when the evidence obtained from individual systems is complementary in nature. Hidden Markov models (HMMs) are based on the maximum likelihood (ML) approach for training CV patterns of variable length. Support vector machine (SVM) models are based on discriminative learning approach for training fixed length CV patterns. Because of the differences in the training methods and in the pattern representation used; they may provide complementary evidence for CV classes. Complementary evidence available from these classifiers is combined using the sum rule. Effectiveness of the multiple classifier system is demonstrated for recognition of CV units of speech in Indian languages.

Analysis of laughter and speech-laugh signals using excitation source information

Speech-laugh is a speech-synchronous form of laughter that often occurs in natural conversation. However, there are deviations in features of speech-laugh when compared with laughter and neutral speech individually. The objective of this study is to analyse the excitation source features to capture the deviations between laughter and speech-laughs in voiced regions. The features used in this analysis are based on instantaneous fundamental frequency and strength of excitation (β) at epochs. Modified zero frequency filtering (ZFF) method is used to extract the features. Kullback-Leibler (KL) distances obtained show that there are deviations in excitation source features which can be exploited to develop a method to discriminate speech-laughs from laughter. Experimental results show that features used are robust and speaker independent in discriminating speech-laughs from laughter. Results showing deviations of laughter and speech-laughs from neutral speech were also presented.

AM-FM model based approach for detection of glottal closure instants

Glottal closure instants (GCI) are the instants of significant excitation of the vocal tract system during the production of speech. The performance of many speech analysis and synthesis approaches depends on accurate estimation of GCIs. A new method for detection of GCI locations in speech utterances using Fourier-Bessel (FB) expansion and amplitude and frequency modulated (AM-FM) signal model is proposed in this paper. The inherent filtering property of the FB expansion is used to weaken the effect of formants in the speech utterances. The band-limited signal reconstructed using FB coefficients has been considered as an AM-FM signal. The discrete-energy separation algorithm (DESA) has been used to estimate the amplitude envelope (AE) function of the AM-FM signal model. The estimated AE function is explored for detection of GCIs. The CMU-Arctic database is used in this work for the validation of the proposed method for GCI detection with the electro-glottograph (EGG) as a reference. It has been observed that the detected GCI by the proposed method seems to provide accurate location of GCIs.

Fourier-Bessel cepstral coefficients for robust speech recognition

In this paper we propose Fourier-Bessel cepstral coefficients (FBCC) features for robust speech recognition. The Fourier-Bessel representation of the speech signal is obtained using Bessel function as a basis set. The FBCC are extracted from zeroth order Bessel coefficients taking into account of the perceptual characteristics of human auditory system. Recognition accuracy is measured using the CMU SPHINX-III speech recognition system using the DARPA Resource Management (RM) speech corpus for training and testing. We evaluate the FBCC in a common experimental set up and compare their performance against traditional technique such as the Mel-frequency cepstral coefficients (MFCC) for various noise conditions. The recognition accuracy is found to be better using FBCC features in comparison with MFCC features under noisy condition data.

Nonlinear principal component analysis for seismic data compression

Seismic data processing to interpret subsurface features is both computationally and data intensive. It is necessary to keep the dimensionality of data as small as possible, for good generalization from limited data. Therefore it is worthwhile exploring methods to compress the size of seismic data. In this paper, we consider approaches for linear and nonlinear principal component analysis (PCA) methods for compression of data for seismic signal processing. Principal component analysis (PCA) can improve seismic interpretations. Linear compression is realized by Karhunen-Loeve transform (KLT) and also by three layer autoassociative neural network (AANN) models. The distribution capturing ability of five layer AANN model is explored for nonlinear principal component analysis for compression of seismic data.

Detection of voice onset time using FB expansion and AM-FM model

The voice onset time (VOT) combines the temporal and frequency structure over very short duration. This makes the VOT detection task difficult. But the VOT is an important temporal feature. In this paper we propose a new method for the detection of VOT in speech utterances. The method uses Fourier-Bessel (FB) expansion followed by amplitude and frequency modulated (AM-FM) signal model. The FB expansion is used to emphasize the vowel and consonant regions of stop consonant vowel (SCV) units (/ka/, /ta/, and /pa/). The emphasized signals using selected range of FB coefficients have been considered narrow-band AM-FM signals. The discrete energy separation algorithm (DESA) has been used to estimate amplitude envelope (AE) function of the AM-FM model. The estimated AE function is explored for the detection of VOT. The VOT detection studies are carried out using the speech corpus consists of utterances of various male and female speakers.