Eliminating inter-speaker variability prior to discriminant transforms

This paper proposes a novel feature-space VTLN (vocal tract length normalization) method that models frequency warping as a linear interpolation of contiguous Mel filter-bank energies. The presented technique aims to reduce the distortion in the Mel filter-bank energy estimation due to the harmonic composition of voiced speech intervals and DFT (discrete Fourier transform) sampling when the central frequency of band-pass filters is shifted. This paper also proposes an analytical maximum likelihood (ML) method to estimate the optimal warping factor in the cepstral space. The presented interpolated filter-bank energy-based VTLN leads to relative reductions in WER (word error rate) as high as 11.2% and 7.6% when compared with the baseline system and standard VTLN, respectively, in a medium-vocabulary continuous speech recognition task. Also, the proposed VTLN scheme can provide significant reductions in WER when compared with state-of-the-art VTLN methods based on linear transforms in the cepstral feature-space. The warping factor estimated with the proposed VTLN approach shows more dependence on the speaker and more independence of the acoustic-phonetic content than the warping factor resulting from standard and state-of-the-art VTLN methods. Finally, the analytical ML-based optimization scheme presented here achieves almost the same reductions in WER as the ML grid search version of the technique with a computational load 20 times lower.

In Vocal Tract Length Normalization (VTLN) a linear or nonlinear frequency transformation compensates for different vocal tract lengths. Finding good estimates for the speaker specific warp parameters is a critical issue. Despite good results using the Maximum Likelihood criterion to find parameters for a linear warping, there are concerns using this method. We searched for a new criterion that enhances the inter-class separability in addition to optimizing the distribution of each phonetic class. Using such a criterion, Linear Discriminant Analysis determines a linear transformation in a lower dimensional space. For VTLN, we keep the dimension constant and warp the training samples of each speaker such that the Linear Discriminant is optimized. Although that criterion depends on all training samples of all speakers it can iteratively provide speaker specific warp factors. We discuss how this approach can be applied in speech recognition and present first results on two different recognition tasks. 1 Speaker Normalization using VTLN Vocal Tract Length Normalization (VTLN) has proven to decrease the word error rate of a speech recognition system, compared to systems not using such an approach to reduce the variability introduced by different speakers. The main effect addressed here is a shift of the formant frequencies of the speakers caused by their different vocal tract lengths. Two issues have been investigated. The first is how to map one speaker’s spectrum on that of a “standard” or average speaker, depending on a warp parameter which is correlated with the vocal tract length. The other issue is how to find an appropriate warp parameter for each speaker. Most studies assume that the same algorithm is used for training and test, but this is not always necessary. [Acero (1990)] has used a bilinear transform with one speaker dependent parameter. In a first attempt he observed that the algorithm chose a degenerate case where all input frames are transformed into a constant. Therefore, he enforced a constant average warping parameter over all speakers. Modeling the vocal tract as a uniform tube of length L, the formant frequencies are proportional to 1/L. Therefore, some approaches use a linear warp of the frequency scale to normalize speakers. The warp can be performed in the time or spectral domain. In the latter case, a new spectrum is derived by interpolation or by modifying the Mel frequency filter bank. When the warp is applied in the spectral domain, the problem of mismatching frequency ranges occurs. [Wegmann et al (1996)] used a piecewise linear spectral mapping to avoid this problem. They estimated the slope of the transformation function based on a maximum likelihood criterion. [Eide and Gish (1996)] proposed a compromise of different vowel models, namely the uniform tube model and the Helmholtz resonator. They warped the frequency axis f of a speaker according to

This work investigates the use of subglottal resonances (SGRs) for speaker normalization in noisy environments. Based on our previous work, a noise-robust algorithm is developed for estimating the first three SGRs from speech signals; it achieves robustness by factoring the short-term (or local) signal-to-noise ratio into the estimation process. The SGR estimates provided by this algorithm are refined by applying maximum-likelihood (ML) corrections, and are used in a non-linear frequency-warping technique that we recently developed. This SGR-based normalization (SN) scheme is evaluated on the AURORA-4 database in clean and noisy conditions. Using power-normalized cepstral coefficients (PNCCs) as front-end features, SN reduces the average word error rate by 8.7% relative to ML-based vocal-tract length normalization (VTLN). A fast version of SN (without ML corrections of SGR estimates) is also found to outperform VTLN (by 5.9% relative); it is computationally less complex than VTLN and hence a potential alternative for real-time applications.

We study the effect of different feature space normalization techniques in adverse acoustic conditions. Recognition tests are reported for cepstral mean and variance normalization, histogram normalization, feature space rotation, and vocal tract length normalization on a German isolated word recognition task with large acoustic mismatch. The training data was recorded in a clean office environment and the test data in cars. Speech recognition failed completely without normalization on the highway dataset, whereas the word error rate could be reduced to 17% using an online setup and to 10% with an offline setup.

Multi-microphone speech recognition integrating beamforming, robust feature extraction, and advanced DNN/RNN backend

Modeling the speaker variability is a key challenge for automatic speech recognition (ASR) systems. In this paper, the learning hidden unit contributions (LHUC) based adaptation techniques with compact speaker dependent (SD) parameters are used to facilitate both speaker adaptive training (SAT) and unsupervised test-time speaker adaptation for end-to-end (E2E) lattice-free MMI (LF-MMI) models. An unsupervised model-based adaptation framework is proposed to estimate the SD parameters in E2E paradigm using LF-MMI and cross entropy (CE) criterions. Various regularization methods of the standard LHUC adaptation, e.g., the Bayesian LHUC (BLHUC) adaptation, are systematically investigated to mitigate the risk of overfitting, on E2E LF-MMI CNN-TDNN and CNN-TDNN- BLSTM models. Lattice-based confidence score estimation is used for adaptation data selection to reduce the supervision label uncertainty. Experiments on the 300-hour Switchboard task suggest that, applying BLHUC in the proposed unsupervised E2E adaptation framework to byte pair encoding (BPE) based E2E LF-MMI systems consistently outperformed the baseline systems by relative word error rate (WER) reductions up to 10.5% and 14.7% on the NIST Hub5’00 and RT03 evaluation sets, and achieved the best performance in WERs of 9.0% and 9.7%, respectively. These results are comparable to the results of state-of-the-art adapted LF-MMI hybrid systems and adapted Conformer-based E2E systems.

It is well known that the recognition performance of an automatic speech recognition (ASR) system is affected by intra-speaker as well inter-speaker variability. The differences in the geometry of vocal organs, pitch and speaking-rate among the speakers are some such inter-speaker variabilities affecting the recognition performance. A mismatch between the training and test data with respect to any of those aforementioned factors leads to increased error rates. An example of acoustically mismatched ASR is the task of transcribing children’s speech on adult data-trained system. A large number of studies have been reported earlier that present a myriad of techniques for addressing acoustic mismatch arising from differences in pitch and dimensions of vocal organs. At the same time, only a few works on speaking-rate adaptation employing timescale modification have been reported. Furthermore, those studies were performed on ASR systems developed using Gaussian mixture models. Motivated by these facts, speaking-rate adaptation is explored in this work in the context of children’s ASR system employing deep neural network-based acoustic modeling. Speaking-rate adaptation is performed by changing the frame-length and overlap during front-end feature extraction process. Significant reductions in errors are noted by speaking-rate adaptation. In addition to that, we have also studied the effect of combining speaking-rate adaptation with vocal-tract length normalization and explicit pitch modification. In both the cases, additive improvements are obtained. To summarize, relative improvements in 15–20% over the baselines are obtained by varying the frame-length and frame-overlap.

An investigation into the use of Bayesian learning of the parameters of a multivariate Gaussian mixture density has been carried out. In a continuous density hidden Markov model (CDHMM) framework, Bayesian learning serves as a unified approach for parameter smoothing, speaker adaptation, speaker clustering, and corrective training. The goal of this study is to enhance model robustness in a CDHMM-based speech recognition system so as to improve performance. Our approach is to use Bayesian learning to incorporate prior knowledge into the CDHMM training process in the form of prior densities of the HMM parameters. The theoretical basis for this procedure is presented and preliminary results applying to HMM parameter smoothing, speaker adaptation, and speaker clustering are given.Performance improvements were observed on tests using the DARPA RM task. For speaker adaptation, under a supervised learning mode with 2 minutes of speaker-specific training data, a 31% reduction in word error rate was obtained compared to speaker-independent results. Using Baysesian learning for HMM parameter smoothing and sex-dependent modeling, a 21% error reduction was observed on the FEB91 test.

The automatic recognition of children's speech in acoustically mismatched conditions is a challenging problem on account of large difference in adults' and children's speech. In literature, this challenge is often addressed through concatenation of various feature/model domain adaptation methods like vocal tract length normalization (VTLN), maximum likelihood linear regression (MLLR) and heteroscedastic linear discriminant analysis (HLDA). But a significant gap in the performance of adults and children still remains. This work explores the eigenvoices (EV) based adaptation for addressing the gap in recognition performance of children's speech on adults' speech trained acoustic models. EV is a fast adaptation approach and helps in an effective gender biasing of the acoustic models. On combining EV with VTLN, MLLR and HLDA, under mismatched condition an absolute improvement of about 50% over the unadapted speaker independent system performance is obtained and thus significantly reducing the gap between the performances for adults and children.

Contextual invariant-integration features for improved speaker-independent speech recognition

A feature extraction and fusion algorithm was constructed by combining principal component analysis (PCA) and linear discriminant analysis (LDA) to detect a fault state of the induction motor. After yielding a feature vector with PCA and LDA from current signal that was measured by an experiment, the reference data were used to produce matching values. In a diagnostic step, two matching values that were obtained by PCA and LDA, respectively, were combined by probability model, and a faulted signal was finally diagnosed. As the proposed diagnosis algorithm brings only merits of PCA and LDA into relief, it shows excellent performance under the noisy environment. The simulation was executed under various noisy conditions in order to demonstrate the suitability of the proposed algorithm and showed more excellent performance than the case just using conventional PCA or LDA.

The performance of automatic speech recognition systems for children’s speech is known to suffer from the large variation and mismatch in the acoustic and linguistic attributes between children’s and adults’ speech. One of the various identified sources of mismatch is the difference in formant frequencies between adults and children. In this paper, we propose a formant modification method to mitigate differences between adults’ and children’s speech and to improve the performance of ASR for children. The explored technique gives a relative 27% improvement in system performance compared to a hybrid DNN-HMM baseline. We also compare the system performance with related speaker adaptation methods like vocal tract length normalization (VTLN) and speaking rate adaptation (SRA) and find that the proposed method gives improvements over them, as well. Combining the proposed method with VTLN and SRA results in a further reduction of WER. We also found that the proposed method performs well even for noisy speech.

This paper discusses the results of the pilot experimental research dedicated to speech recognition and perception of the semantic content of the utterances in noisy environment. The experiment included perceptual-auditory analysis of words and phrases in Russian and German (in comparison) in the same noisy environment: various (pink and white) types of noise with various levels of signal-to-noise ratio. The statistical analysis showed that intelligibility and perception of the speech in noisy environment are influenced not only by noise type and its signal-to-noise ratio, but also by some linguistic and extralinguistic factors, such as the existing redundancy of a particular language at various levels of linguistic structure, changes in the acoustic characteristics of the speaker while switching from one language to another one, the level of speaker and listener’s proficiency in a specific language, and acoustic characteristics of the speaker’s voice.

Despite the importance of data reduction as part of the processing of reflection-based classifications, this study represents one of the first in which the effects of both spatial and spectral data reductions on classification accuracies are quantified. Furthermore, the effects of approaches to data reduction were quantified for two separate classification methods, linear discriminant analysis (LDA) and support vector machine (SVM). As the model dataset, reflection data were acquired using a hyperspectral camera in 230 spectral channels from 401 to 879 nm (spectral resolution of 2.1 nm) from field pea (Pisum sativum) samples with and without internal pea weevil (Bruchus pisorum) infestation. We deployed five levels of spatial data reduction (binning) and eight levels of spectral data reduction (40 datasets). Forward stepwise LDA was used to select and include only spectral channels contributing the most to the separation of pixels from non-infested and infested field peas. Classification accuracies obtained with LDA and SVM were based on the classification of independent validation datasets. Overall, SVMs had significantly higher classification accuracies than LDAs (P < 0.01). There was a negative association between pixel resolution and classification accuracy, while spectral binning equivalent to up to 98% data reduction had negligible effect on classification accuracies. This study supports the potential use of reflection-based technologies in the quality control of food products with internal defects, and it highlights that spatial and spectral data reductions can (1) improve classification accuracies, (2) vastly decrease computer constraints, and (3) reduce analytical concerns associated with classifications of large and high-dimensional datasets.

A brain-computer interface (BCI) is a system that enables control of devices or communication with other persons, only through cerebral activity, without using muscles. The main application for BCIs is assistive technology for disabled persons. Examples for devices that can be controlled by BCIs are artificial limbs, spelling devices, or environment control systems. BCI research has seen renewed interest in recent years, and it has been convincingly shown that communication via a BCI is in principle feasible. However, present day systems still have shortcomings that prevent their widespread application. In part, these shortcomings are caused by limitations in the functionality of the pattern recognition algorithms used for discriminating brain signals in BCIs. Moreover, BCIs are often tested exclusively with able-bodied persons instead of conducting tests with the target user group, namely disabled persons. The goal of this thesis is to extend the functionality of pattern recognition algorithms for BCI systems and to move towards systems that are helpful for disabled users. We discuss extensions of linear discriminant analysis (LDA), which is a simple but efficient method for pattern recognition. In particular, a framework from Bayesian machine learning, the so-called evidence framework, is applied to LDA. An algorithm is obtained that learns classifiers quickly, robustly, and fully automatically. An extension of this algorithm allows to automatically reduce the number of sensors needed for acquisition of brain signals. More specifically, the algorithm allows to perform electrode selection. The algorithm for electrode selection is based on a concept known as automatic relevance determination (ARD) in Bayesian machine learning. The last part of the algorithmic development in this thesis concerns methods for computing accurate estimates of class probabilities in LDA-like classifiers. These probabilities are used to build a BCI that dynamically adapts the amount of acquired data, so that a preset, approximate bound on the probability of misclassifications is not exceeded. To test the algorithms described in this thesis, a BCI specifically tailored for disabled persons is introduced. The system uses electroencephalogram (EEG) signals and is based on the P300 evoked potential. Datasets recorded from five disabled and four able-bodied subjects are used to show that the Bayesian version of LDA outperforms plain LDA in terms of classification accuracy. Also, the impact of different static electrode configurations on classification accuracy is tested. In addition, experiments with the same datasets demonstrate that the algorithm for electrode selection is computationally efficient, yields physiologically plausible results, and improves classification accuracy over static electrode configurations. The classification accuracy is further improved by dynamically adapting the amount of acquired data. Besides the datasets recorded from disabled and able-bodied subjects, benchmark datasets from BCI competitions are used to show that the algorithms discussed in this thesis are competitive with state-of-the-art electroencephalogram (EEG) classification algorithms. While the experiments in this thesis are uniquely performed with P300 datasets, the presented algorithms might also be useful for other types of BCI systems based on the EEG. This is the case because functionalities such as robust and automatic computation of classifiers, electrode selection, and estimation of class probabilities are useful in many BCI systems. Seen from a more general point of view, many applications that rely on the classification of cerebral activity could possibly benefit from the methods developed in this thesis. Among the potential applications are interrogative polygraphy ("lie detection") and clinical applications, for example coma outcome prognosis and depth of anesthesia monitoring.