Discriminative feature selection for speech recognition

doi:10.1006/csla.1993.1012

Computer Speech & Language

Volume 7, Issue 3, July 1993, Pages 229-246

https://doi.org/10.1006/csla.1993.1012 Get rights and content

Abstract

Over the last several years, a major factor in reducing the error rate on most speech recognition systems has been the addition of new feature components to the frame vectors. However, because of the larger dimensionality of the frame feature vector, the number of model parameters and the computational requirements have also increased. To improve recognition performance, it is not feasible to indefinitely increase the size of the frame feature vectors, nor is it satisfactory or practical to run experiments on combinatorially chosen subsets of the feature set to pick the best performing subspace with the desired number of dimensions. It becomes clearly desirable to understand which components of the frame feature vector provide the greatest contribution to recognition performance and to discard the least useful components. Our feature ordering method allows new sets of features to be selectively incorporated into existing signal analysis methods.

Discriminative analysis has been successfully used for hidden Markov model (HMM) parameter estimation. In this study, we use discriminative methods to perform feature selection of the frame feature space. The components of the feature vectors are rank ordered according to an objective criterion and only the most "significant" are used for recognition. The proposed feature reduction method has been applied to a 38 dimension vector consisting of 1st and 2nd order time derivatives of the frame energy and of the cepstral coefficients with their 1st and 2nd derivatives. Speaker independent recognition experiments with reduced feature sets were performed on three databases of high quality non-telephone speech and a data base of telephone-based speech recorded during a field trial. The dimension of the frame feature vectors, and hence the number of model parameters, were greatly reduced (by a factor of 2) without a significant loss of recognition performance.

References (0)

Cited by (27)

Feature selection for reduced-bandwidth distributed speech recognition
2012, Speech Communication
Citation Excerpt :
In a multi-speaker isolated digit recognition task the results indicated that it is possible to reduce the feature vector size without impacting the recognition performance. Bocchieri and Wilpon (1993) use discriminative analysis to select a subset of coefficients from the feature vectors in continuous speech recognition experiments. Nicholson et al. (1997) measured the correlation between MFCC feature sets that have good class discrimination and those that produced good speech recognition results.
The impact on speech recognition performance in a distributed speech recognition (DSR) environment of two methods used to reduce the dimension of the feature vectors is examined in this paper. The motivation behind reducing the dimension of the feature set is to reduce the bandwidth required to send the feature vectors over a channel from the client front-end to the server back-end in a DSR system. In the first approach, the features are empirically chosen to maximise recognition performance. A data-centric transform-based dimensionality-reduction technique is applied in the second case. Test results for the empirical approach show that individual coefficients have different impacts on the speech recognition performance, and that certain coefficients should always be present in an empirically selected reduced feature set for given training and test conditions. Initial results show that for the empirical method, the number of elements in a feature vector produced by an established DSR front-end can be reduced by 23% with low impact on the recognition performance (less than 8% relative performance drop compared to the full bandwidth case). Using the transform-based approach, for a similar impact on recognition performance, the number of feature vector elements can be reduced by 30%. Furthermore, for best recognition performance, the results indicate that the SNR of the speech signal should be considered using either approach when selecting the feature vector elements that are to be included in a reduced feature set.
A two level strategy for audio segmentation
2011, Digital Signal Processing: A Review Journal
In this paper we are dealing with audio segmentation. The audio tracks are sampled in short sequences which are classified into several classes. Every sequence can then be further analyzed depending on the class it belongs to. We first describe simple techniques for segmentation in two or three classes. These methods rely on amplitude, spectral or cepstral analysis, and classical Hidden Markov Models. From the limitations of these approaches, we propose a two level segmentation process. The segmentation is performed by computing several features for each audio sequence. These features are computed either on a complete audio segment or on a frame (set of samples) which is a subset of the audio segment. The proposed approach for microsegmentation of audio data consists of a combination of a K-mean classifier at the segment level and of a Multidimensional Hidden Markov Model system using the frame decomposition of the signal. A first classification is obtained using the K-mean classifier and segment-based features. Then final result comes from the use of Multidimensional Hidden Markov Models and frame-based features involving temporary results. Multidimensional Hidden Markov Models are an extension of classical Hidden Markov Models dedicated to multicomponent data. They are particularly adapted to our case where each audio segment can be characterized by several features of different natures. We illustrate our methods in the context of analysis of football audio tracks.
Use of voicing features in HMM-based speech recognition
2002, Speech Communication
We investigate speech recognition features related to voicing functions that indicate whether the vocal folds are vibrating. We describe two voicing features, periodicity and jitter, and demonstrate that they are powerful voicing discriminators. The periodicity and jitter features and their first and second time derivatives are appended to a standard 38-dimensional feature vector comprising the first and second time derivatives of the frame energy and the cepstral coefficients with their first and second time derivatives. HMM-based connected-digit (CD) and large-vocabulary (LV) recognition experiments comparing the traditional and extended feature sets show that voicing features and spectral information are complementary and that improved speech recognition performance is obtained by combining the two sources of information. We further conclude that the difference in performance with and without voicing becomes more significant when minimum string error (MSE) training is used than when maximum likelihood (ML) training is used.
Combining acoustic and articulatory feature information for robust speech recognition
2002, Speech Communication
Citation Excerpt :
The MFCCs which were eliminated in favor of these AFs are the first derivative of the 12th cepstral coefficient and the second derivatives of the 4th, 6th, 7th, 9th, 11th and 12th cepstral coefficients. This result confirms the low relevance of delta–delta coefficients observed previously (Bocchieri and Wilpon, 1993) and the importance of the place of articulation dimension generally acknowledged in phonetics. The resulting 39-dimensional feature set was used to train another recognition system with a 256-class full-covariance codebook.
The idea of using articulatory representations for automatic speech recognition (ASR) continues to attract much attention in the speech community. Representations which are grouped under the label “articulatory” include articulatory parameters derived by means of acoustic-articulatory transformations (inverse filtering), direct physical measurements or classification scores for pseudo-articulatory features. In this study, we revisit the use of features belonging to the third category. In particular, we concentrate on the potential benefits of pseudo-articulatory features in adverse acoustic environments and on their combination with standard acoustic features. Systems based on articulatory features only and combined acoustic-articulatory systems are tested on two different recognition tasks: telephone-speech continuous numbers recognition and conversational speech recognition. We show that articulatory feature (AF) systems are capable of achieving a superior performance at high noise levels and that the combination of acoustic and AFs consistently leads to a significant reduction of word error rate across all acoustic conditions.
Die Idee, artikulatorische Repräsentationen zur automatischen Spracherkennung zu nutzen, erweckt auch weiterhin großes Interesse in der Sprachverarbeitungsforschung. Repräsentationen, die unter dem Schlagwort “artikulatorisch” zusammengefaßt werden, umfassen artikulatorische Parameter, die mit Hilfe von akustisch-artikulatorischen Transformationen (inverser Filterung) erzeugt werden, direkte physikalische Messwerte oder Klassifikationsbewertungen für pseudo-artiulatorische Merkmale. In dieser Arbeit untersuchen wir die Verwendung von Merkmalen der letzteren Kategorie. Speziell konzentrieren wir uns dabei auf die möglichen Vorteile pseudo-artikulatorischer Merkmale unter ungünstigen akustischen Bedingungen und auf ihre Kombination mit herkömmlichen akustischen Merkmalen. Systeme, die auf artikulatorischen Merkmale allein basieren, und kombinierte artikulatorisch-akustische Systeme werden auf zwei unterschiedlichen Erkennungsaufgaben evaluiert: der Erkennung von Zahlenfolgen in Telephonqualität sowie der Erkennung spontan gesprochener Sprache. Wir zeigen, daß durch die Verwendung artikulatorischer Merkmale eine Verbesserung der Leistungsfähigkeit bei hohen Geräuschpegeln erreicht wird, und daß die Kombination von akustischen und artikulatorischen Merkmalen konsistent zu einer signifikanten Reduktion der Fehlerrate unter allen akustischen Bedingungen führt.
Heteroscedastic discriminant analysis and reduced rank HMMs for improved speech recognition
1998, Speech Communication
We present the theory for heteroscedastic discriminant analysis (HDA), a model-based generalization of linear discriminant analysis (LDA) derived in the maximum-likelihood framework to handle heteroscedastic-unequal variance-classifier models. We show how to estimate the heteroscedastic Gaussian model parameters jointly with the dimensionality reducing transform, using the EM algorithm. In doing so, we alleviate the need for an a priori ad hoc class assignment. We apply the theoretical results to the problem of speech recognition and observe word-error reduction in systems that employed both diagonal and full covariance heteroscedastic Gaussian models tested on the TI-DIGITS database.
Isolated word recognition using modular recurrent neural networks
1998, Pattern Recognition
This paper describes a novel method of using recurrent neural networks (RNN) for isolated word recognition. Each word in the target vocabulary is modeled by a fully connected recurrent network. To recognize an input utterance, the best matching word is determined based on its temporal output response. The system is trained in two stages. First, the RNN speech models (RSM) are trained independently to capture the essential static and temporal characteristics of individual words. This is performed by using an iterative re-segmentation training algorithm which gives the optimal phonetic segmentation automatically for each training utterance. The second-stage involves mutually discriminative training among the RSMs, aiming at minimizing the probability of misclassification. A series of simulation experiments have been performed to demonstrate the effectiveness of the proposed recognition method. For the recognition of (A) 20 English words, (B) 11 Cantonese digits and (C) 58 Cantonese CV syllables, the top-1 accuracy are 91.9, 93.6 and 87.1%, respectively.

View all citing articles on Scopus

View full text

Regular ArticleDiscriminative feature selection for speech recognition

Abstract

Regular Article
Discriminative feature selection for speech recognition