In this paper, we introduce a novel method for clustering speech gestures, represented as continuous trajectories in acoustic parameter space. Trajectory Clustering allows us to avoid the conditional independence assumption that makes it difficult to account for the fact that successive measurements of an articulatory gesture are correlated. We apply the trajectory clustering method for developing multiple parallel hidden Markov models (HMMs) for a continuous digits recognition task. We compare the performance obtained with data-driven clustering to the recognition performance obtained with conventional head-body-tail models, which use knowledge-based criteria for building multiple HMMs in order to obviate the trajectory folding problem. The results show that trajectory clustering is able to discover structure in the the training database that is different from the structure assumed by the knowledge-based approach. In addition, the data-derived structure gives rise to significantly better recognition performance, and results in a 10% word error rate reduction.

Contains fulltext : 42030.pdf (Publisher’s version ) (Closed access)

Making ASR noise robust requires a form of data normalisation to ensure that the distributions of acoustic features in the training and test condition look similar. Usually, it is attempted to compensate for the impact of noise by estimating the noise characteristics from the signal. In this paper we explore a new method that builds on a priori knowledge stored in clean speech models. Using Mel bank log-energy features, classical clean speech HMMs were replaced by models in which the model components corresponding to low energy are not considered during recognition. Application of the new method to clean matched data showed that recognition performance was equal or better compared to baseline when less than 45% of the model components were discarded. In the case of noisy data, the performance gains were marginal for the model component selections studied so far. Analysis of the results suggests that future research should focus on combining the new model-driven approach with data-driven methods.

Within the Aurora2 experimental framework, the aim of this study is to determine what the relative contributions of spectral shape and energy features are to the mismatch observed between clean training and noisy test data. In addition to measurements on the baseline Aurora2 system, recognition performance was also evaluated after the application of time domain noise reduction (TDNR) and histogram normalisation (HN) in the cepstral domain. The results indicate that, for the Aurora2 digit recognition task, TDNR, HN, as well as a combination of the two techniques achieve higher recognition rates by reducing the mismatch in the energy part of acoustic feature space. The corresponding mismatch reduction in the spectral shape features yields hardly any gain in recognition performance.

The current version of the psycholinguistic model of human word recognition Shortlist suffers from two unrealistic constraints. First, the input of Shortlist must consist of a single string of phoneme symbols. Second, the current version of the search in Shortlist makes it difficult to deal with insertions and deletions in the input phoneme string. This research attempts to fully automatically derive a phoneme string from the acoustic signal that is as close as possible to the number of phonemes in the lexical representation of the word. We optimised an Automatic Phone Recogniser (APR) using two approaches, viz. varying the value of the mismatch parameter and optimising the APR output strings on the output of Shortlist. The approaches show that it will be very difficult to satisfy the input requirements of the present version of Shortlist with a phoneme string generated by an APR.

In this paper, the accumulated Kullback divergence (AKD) is used to analyze ASR performance deterioration due to the presence of background noise. The AKD represents a distance between the feature value distribution observed during training and the distribution of the observations in the noisy test condition for each individual feature vector component. In our experiments the AKD summed over all feature vector components shows a high correlation with word error rate and AKD computed per component can be used to pinpoint those feature vector components that substantially contribute to recognition errors. It is argued that the distance measure could be a useful evaluation tool for analyzing the strengths and weaknesses of existing noise robustness approaches and might help to suggest research strategies that focus on those elements of the acoustic feature vector that are most severely affected by the noise.

This paper describes the noise robust feature extraction meth ods developed by France Telecom and Alcatel for the noise robust front-end standardisation of ETSI Aurora.It is shown that both noise reduction methods give a substantial im provement when compared to a standard MFCC feature ex traction algorithm for speech recognition in noisy environ ments.In addition, blind equalisation and feature vector se lection were used for further improvement of recognition performance.Results are discussed for the ETSI Aurora 2 task and the SDC-Italian task as well.It was found that the combi nation of noise reduction with the proposed methods is capa ble to achieve around 50% reduction of the error rate.In the context of the open ETSI Aurora standardisation, two propos als were submitted based on these methods, they achieved the best results among all the proposals.

\n Contains fulltext :\n 75051.pdf (author's version ) (Open Access)\n

Within the context of robust acoustic features for automatic speech recognition (ASR), we evaluated mel-frequency cepstral coefficients (MFCCs) derived from two spectral representation techniques, i.e. the fast Fourier transform (FFT) and linear predictive coding (LPC). ASR systems based on the two feature types were tested on a digit recognition task using continuous density hidden Markov phone models. System performance was determined in clean acoustic conditions as well as in different simulations of adverse acoustic conditions. The LPC-based MFCCs outperformed their FFT counterparts in most of the adverse acoustic conditions that were investigated in this study. A tentative explanation for this difference in recognition performance is given.

Within the context of automatic speech recognition (ASR) applications for telephony, we investigate the acoustic preprocessing issues that are at stake in going from the fixed line to the cellular network. Because the spectral representation used in enhanced full rate GSM is linear prediction, we investigate the relative advantages and drawbacks of conventional mel-frequency cepstral coefficient (MFCC) parameters derived from a non-parametric fast Fourier transform (FFT) and MFCC parameters derived from a linear predictive coding (LPC) spectral estimate. Robust formant parameters, also derived from an LPC description of the spectrum, are studied as an alternative to MFCCs. Within the framework of connected digit recognition based on hidden Markov models, ASR performance was measured for clean conditions, as well as for three different additive noise conditions. In addition, the performance of a conventional recognition procedure was compared with the performance of an ASR system based on our acoustic backing-off implementation of missing feature theory (MFT).

In this paper we investigate acoustic backing-off as an operationalization of Missing Feature Theory with the aim to increase recognition robustness. Acoustic backing-off effectively diminishes the detrimental influence of outlier values by using a new model of the probability density function of the feature values. The technique avoids the need for explicit outlier detection. Situations that are handled best by Missing Feature Theory are those where only part of the coefficients are disturbed and the rest of the vector is unaffected. Consequently, one may predict that acoustic feature representations that smear local spectrotemporal distortions over all feature vector elements are inherently less suitable for automatic speech recognition. Our experiments seem to confirm this prediction. Using additive band limited noise as a distortion and comparing four different types of feature representations, we found that the best recognition performance is obtained with recognizers that use acoustic backingoff and that operate on feature types that minimally smear the distortion.

In this paper we investigate the use of broad phonetic class (BPC) models in a text independent speaker recognition task. These models can be used to bring down the variability due to the intrinsic differences between mutual phonetic classes in the speech material used for training of the speaker models. Combining BPC recognition with text independent speaker recognition moves a bit in the direction of text dependent speaker recognition: a task which is known to reach better performance. The performance of BPC modelling is compared to our baseline system using ergodic 5-state HMMs. The question which BPC contains most speaker specific information is addressed. Also, it is investigated if and how the BPC alignment is correlated with the state alignment from the baseline system to check the assumption that states of an ergodic HMM can model broad phonetic classes [3]. 1. Introduction For text-independent speaker recognition HMM-based methods have been popular for several years now. Espe...

Phase-corrected RASTA is a new technique for channel normalization that consists of classical RASTA filtering followed by a phase correction operation. In this manner, the channel bias is as effectively removed as with classical RASTA, without introducing a left context dependency. The performance of the phase-corrected RASTA channel normalization technique was evaluated for a continuous speech recognition task. Using context-independent hidden Markov models we found that phase-corrected RASTA reduces the best-sentence word error rate (WER) by 23% compared to classical RASTA. For contextdependent models phase-corrected RASTA reduces WER by 15% compared to classical RASTA. 1. INTRODUCTION In order to reduce the linear filtering effect of communication channels, different channel normalization (CN) techniques have been proposed (e.g. [1, 2, 3]). Recently, a new, extended version of the classical RASTA filtering technique was proposed and tested in the context of connected digit recogn...

In this paper we propose to introduce backing-off in the acoustic contributions of the local distance functions used during Viterbi decoding as an operationalisation of missing feature theory for increased recognition robustness. Acoustic backing-off effectively removes the detrimental influence of outlier values from the local decisions in the Viterbi algorithm. It does so without the need for prior knowledge that specific features are missing. Acoustic backing-off avoids any kind of explicit outlier detection. This paper provides a proof of concept of acoustic backing-off in the context of connected digit recognition over the telephone, using artificial distortions of the acoustic observations. It is shown that the word error rate can be maintained at the level of 2:5% obtained for undisturbed features, even in the case where a conventional local distance computation without backing-off leads to a word error rate ? 80:0%. The approach appears to be able to handle up to four independe...

In this paper we propose an extension to the classical RASTA technique. The new method consists of classical RASTA filtering followed by an all-pass phase correction filter. In this manner, the influence of the communication channel is as effectively removed as with classical RASTA. However, our proposal does not introduce a left-context dependency like classical RASTA. Therefore the new method is better suited for automatic speech recognition based on context-independent modeling with Gaussian mixture hidden Markov models. We tested this in the context of connected digit recognition over the phone. In case we used context-dependent hidden Markov models (i.e., word models), we found that classical RASTA and phase-corrected RASTA performed equally well. For context-independent phone-based models, we found that phase-corrected RASTA can outperform classical RASTA depending on the acoustic resolution of the models.

We compared three different channel normalisation (CN) meth ods in the context of a connected digit recognition task over the phone: ceptrum mean substraction (CMS), RASTA filtering and the Gaussian dynamic cepstrum reprsentation (GDCR).Using a small set of context-independent (CI) continuous Gaussian mixture hid den Markov models (HMMs) we found that CMS and RASTA out performed the GDCR technique.We show that the main cause for the superiority of CMS compared to RASTA is the phase distortion introduced by the RASTA filter.Recognition results for a phasecorrected RASTA technique are identical to those of CMS.Our re sults indicate that an ideal cepstrum based CN method should (1) effectively remove the DC-component, (2) at least preserve modu lation frequencies in the range 2-16 Hz and (3) introduce no phase distortion in case CI HMMs are used for recognition.

The experiments described show that RARMA (robust autoregressive moving average) analysis can be applied to the extraction of speech synthesis rules for nasal consonants that bear relationship to articulatory gestures. This is an important advantage in comparison to the results of the usual all-pole techniques. The poles and zeros resolved by RARMA could be directly translated into target frequencies for the vocal tract (anti-)resonators in this synthesis model. Inspection of short-term FFTs of natural speech utterances showed that during the realization of the nasal consonant an additional zero-per-pole is present around 800 Hz that is not resolved by RARMA. This special event rides on the steep spectral slope of the first resonance peak in the nasal consonant.>

Two methods for deriving synthesizer control rules are discussed. In one approach a trained phonetician interprets the results of a robust autoregressive moving average (RARMA) analysis and makes a separation between spectral events attributable to the source and the vocal tract, respectively, in order to derive rules for controlling the poles and zeros in the tract model. This procedure appears to be fruitful but rather time-consuming. In the other approach an attempt is made to support the procedure as much as possible by automatic enhancement of the analysis results. The experiments indicate that RARMA analysis delivers reliable pole and zero tracks that can be associated with dynamic characteristics of the vocal tract.>

In order to derive voice source control rules from natural speech, the parameters of a source model must be derived from the acoustic signal. This is done by parameterizing the results of glottal inverse filtering. A number of different inverse filtering procedures and the ease with which their results can be parameterized are compared. It is shown that closed glottic interval covariance linear predictive coding is as powerful as more sophisticated techniques, because it is the only known method that can strictly be limited to the closed glottis interval.>