Test Speech Data Research Articles

PaSCoNT - Parallel Speech Corpus of Northern-central Thai for automatic speech recognition

This paper proposed a Parallel Speech Corpus of Northern-central Thai (PaSCoNT). The purpose of this research is not only to understand the different linguistic characteristics between Northern and Central Thai, but also to utilize this corpus for automatic speech recognition. The corpus is composed of speech data from dialogues of daily life among northern Thai people. We designed 2,000 Northern Thai sentences covering all phonemes, in collaboration with linguists specialized in the Northern Thai dialect. The samples in this study are 200 Northern Thai dialect speakers who had been living in Chiang Mai province for more than 18 years. The speech was recorded in both open and closed environments. In the speech recording, each speaker must read 100 pairs of Northern-Central Thai sentences to ensure that the speech data comes from the same speaker. In total, 100 h of speech were recorded: 50 h of Northern Thai and 50 h of Central Thai. Overall, PaSCoNT consists of 907,832 words and 6,279 vocabulary items. Statistical analysis of the PaSCoNT corpus revealed that 49.64 % of words in the lexicon belongs to the Northern Thai dialect, 50.36 % from the Central Thai dialect, and 1,621 vocabulary items appeared in both Northern and Central Thai. Statistical analysis is used to examine the difference in speech tempo, i.e. time per phoneme (TTP), syllable per minute (SPM), between Northern and Central Thai. The results revealed that there were statistically significant differences speech tempo between Central and Northern Thai. The TTP speaking and articulation rate of Central Thai is lower than Northern Thai whereas SPM speaking and articulation rate of Central Thai is higher than Northern Thai. The results also showed that the ASR model training using Northern Thai speech corpus provides the lower WER% when testing using Northern Thai testing speech data and provides the higher WER% when testing using Central Thai Testing speech data and vice versa. However, the ASR model training using the PaSCoNT speech corpus provides the lower WER% for both Northern Thai and Central Thai testing speech data.

Fusion Multistyle Training for Speaker Identification of Disguised Speech

Determining the speaker of a given speech utterance from a group of people is referred to as speaker identification. When voice disguising is done by a person, which is commonly seen in crime scenes, a mismatch between the training and the test speech data occurs, referred to as mismatched problem. It markedly decreases the performance of the speaker identification system. To address this mismatched problem, various multistyle training strategies and a fusion method were previously studied by the authors. This paper further investigates the performance of three multiple-model methods at the decision level for this mismatched problem and compare its performance with the previously studied multistyle training strategies. It is found that the fusion of the two multistyle training strategies, outperformed all other single style training and the multiple-model methods investigated on an average across the different test speech data. This fusion multistyle training technique can be easily employed in a security conscious organization, where monitoring of the employees are required.

Explicit Pitch Mapping for Improved Children’s Speech Recognition

Recognizing children’s speech on automatic speech recognition (ASR) systems developed using adults’ speech is a very challenging task. As reported by several earlier works, a severely degraded recognition performance is observed in such ASR tasks. This is mainly due to the gross mismatch in the acoustic and linguistic attributes between those two groups of speakers. One among the various identified sources of mismatch is that the vocal organs of the adult and child speakers are of significantly different dimensions. Feature-space normalization techniques are noted to effectively address the ill-effects arising from those differences. Two most commonly used approaches are the vocal-tract length normalization and the feature-space maximum-likelihood linear regression. Another important mismatch factor is the large variation in the average pitch values across the adult and child speakers. Addressing the ill-effects introduced by the pitch differences is the primary focus of the presented study. In this regard, we have explored the feasibility of explicitly changing the pitch of the children’s speech so that observed pitch differences between the two groups of speaker are reduced. In general, speech data from children is high-pitched in comparison with that from the adults’. Consequently, in this study, the pitch of the adults’ speech used for training the ASR system is kept unchanged while that for the children’s test speech data is reduced. Significant improvement in the recognition performance is noted by this explicit reduction of pitch. To conserve the critical spectral information and to avoid introducing perceptual artifacts, we have exploited timescale modification techniques for explicit pitch mapping. Furthermore, we also presented two schemes to automatically determine the factor by which the pitch of the given test data should be varied. Automatically determining the compensation factor is critical since an ASR system is expected to be accessed by both adult and child speakers. The effectiveness of proposed techniques is evaluated on adult data trained ASR systems employing different acoustic modeling approaches, viz. Gaussian mixture modeling (GMM), subspace GMM and deep neural networks (DNN). The proposed techniques are found to be highly effective in all the explored modeling paradigms. To further study the effectiveness of the proposed approaches, another DNN-based ASR system is developed on a mix of speech data from adult as well as child speakers. The use of pitch reduction is observed to be effective even in this case.

Experimental Study on Extreme Learning Machine Applications for Speech Enhancement

In wireless telephony and audio data mining applications, it is desirable that noise suppression can be made robust against changing noise conditions and operates in real time (or faster). The learning effectiveness and speed of artificial neural networks are therefore critical factors in applications for speech enhancement tasks. To address these issues, we present an extreme learning machine (ELM) framework, aimed at the effective and fast removal of background noise from a single-channel speech signal, based on a set of randomly chosen hidden units and analytically determined output weights. Because feature learning with shallow ELM may not be effective for natural signals, such as speech, even with a large number of hidden nodes, hierarchical ELM (H-ELM) architectures are deployed by leveraging sparse autoencoders. In this manner, we not only keep all the advantages of deep models in approximating complicated functions and maintaining strong regression capabilities, but we also overcome the cumbersome and time-consuming features of both greedy layer-wise pre-training and back-propagation (BP)-based fine tuning schemes, which are typically adopted for training deep neural architectures. The proposed ELM framework was evaluated on the Aurora-4 speech database. The Aurora-4 task provides relatively limited training data, and test speech data corrupted with both additive noise and convolutive distortions for matched and mismatched channels and signal-to-noise ratio (SNR) conditions. In addition, the task includes a subset of testing data involving noise types and SNR levels that are not seen in the training data. The experimental results indicate that when the amount of training data is limited, both ELMand H-ELM-based speech enhancement techniques consistently outperform the conventional BP-based shallow and deep learning algorithms, in terms of standardized objective evaluations, under various testing conditions.

Noise-robust speech feature processing with empirical mode decomposition

Abstract In this article, a novel technique based on the empirical mode decomposition methodology for processing speech features is proposed and investigated. The empirical mode decomposition generalizes the Fourier analysis. It decomposes a signal as the sum of intrinsic mode functions. In this study, we implement an iterative algorithm to find the intrinsic mode functions for any given signal. We design a novel speech feature post-processing method based on the extracted intrinsic mode functions to achieve noise-robustness for automatic speech recognition. Evaluation results on the noisy-digit Aurora 2.0 database show that our method leads to significant performance improvement. The relative improvement over the baseline features increases from 24.0 to 41.1% when the proposed post-processing method is applied on mean-variance normalized speech features. The proposed method also improves over the performance achieved by a very noise-robust frontend when the test speech data are highly mismatched.

Compensating Acoustic Mismatch Using Class-Based Histogram Equalization for Robust Speech Recognition

A new class-based histogram equalization method is proposed for robust speech recognition. The proposed method aims at not only compensating for an acoustic mismatch between training and test environments but also reducing the two fundamental limitations of the conventional histogram equalization method, the discrepancy between the phonetic distributions of training and test speech data, and the nonmonotonic transformation caused by the acoustic mismatch. The algorithm employs multiple class-specific reference and test cumulative distribution functions, classifies noisy test features into their corresponding classes, and equalizes the features by using their corresponding class reference and test distributions. The minimum mean-square error log-spectral amplitude (MMSE-LSA)-based speech enhancement is added just prior to the baseline feature extraction to reduce the corruption by additive noise. The experiments on the Aurora2 database proved the effectiveness of the proposed method by reducing relative errors by over the mel-cepstral-based features and by over the conventional histogram equalization method, respectively.

Class-Based Histogram Equalization for Robust Speech Recognition

A new class-based histogram equalization method is proposed for robust speech recognition. The proposed method aims at not only compensating the acoustic mismatch between training and test environments, but also at reducing the discrepancy between the phonetic distributions of training and test speech data. The algorithm utilizes multiple class-specific reference and test cumulative distribution functions, classifies the noisy test features into their corresponding classes, and equalizes the features by using their corresponding class-specific reference and test distributions. Experiments on the Aurora 2 database proved the effectiveness of the proposed method by reducing relative errors by 18.74%, 17.52%, and 23.45% over the conventional histogram equalization method and by 59.43%, 66.00%, and 50.50% over mel-cepstral-based features for test sets A, B, and C, respectively.

Method for speech processing involving whole-utterance modeling

A speech verification process involves comparison of enrollment and test speech data and an improved method of comparing the data is disclosed, wherein segmented frames of speech are analyzed jointly, rather than independently. The enrollment and test speech are both subjected to a feature extraction process to derive fixed-length feature vectors, and the feature vectors are compared, using a linear discriminant analysis and having no dependence upon the order of the words spoken or the speaking rate. The discriminant analysis is made possible, despite a relatively high dimensionality of the feature vectors, by a mathematical procedure provided for finding an eigenvector to simultaneously diagonalize the between-speaker and between-channel covariances of the enrollment and test data.

A 6.7 kbps vector sum excited linear prediction on TMS320C54X digital signal processor

In this paper, a 6.7-kbps vector sum excited linear prediction (VSELP) coder with less computational complexity is presented. A very efficient VSELP codebook with nine basis vectors and a heuristic K-selection method (to reduce the search space and complexity) is constructed to obtain the stochastic codebook vector. The nine basis vectors are obtained by optimizing a set of randomly generated basis vectors. During the optimization process, we have trained the basis vectors to give the system apriori knowledge of the characteristics of the input. The coder is implemented on a TMS320C541 digital signal processor. The performance is evaluated by testing the 6.7-kbps VSELP coder with different test speech data taken from different speakers. The quality of the coder is estimated by comparing the performance of the 6.7-kbps VSELP coder with an 8-kbps VSELP speech coder based on the IS-54 standards.

Self-learning speaker and channel adaptation based on spectral variation source decomposition

A self-learning speaker and channel adaptation technique based on the separation of speech spectral variation sources is developed for improving speaker-independent continuous speech recognition. Statistical methods are formulated to remove spectral biases at the acoustic level and to adapt parameters of Gaussian mixture densities at the phone unit level. The spectral bias is estimated in two steps using unsupervised maximum likelihood estimation: the probability distributions of the speech spectral features are first assumed uniform for severely mismatched channels, and the spectral bias is then reestimated using Gaussian phone models. Unsupervised sequential phone model adaptation (USPA) is performed via Bayesian estimation from the on-line, bias-removed speech data, and iterative phone model adaptation (IPA) is further performed for dictation applications. The task vocabulary size was 853; the grammar perplexity was 105; the test speech data were collected under mismatched recording conditions with each test set containing 198 sentences. Depending on the speakers and channel conditions, the two-step spectral bias removal yielded relative error reductions (RER) of 3% to 11% compared to the conventional cepstral mean removal; the USPA yielded RER of 12% to 26% after the two-step bias removal; the IPA further yielded RER of 8% to 19% after the USPA.

A unified approach to speaker verification with noisy speech inputs

Several speaker verification techniques based on different features of the human voice and with verification accuracy better than 99% have been proposed in the literature. Most of these systems work well only when the reference and test speech utterances are recorded under the same relatively noise-free conditions. The goal of this study was to develop a speaker verification system based on an orthogonal linear prediction model, that adopts a unified approach by utilizing only one set of reference parameters irrespective of whether the test speech data is of high quality or noise corrupted. Initial investigations reveal that significant improvement in accuracy is achieved if the verification algorithm utilizes only those parameters that are least sensitive to additive broadband noise. However, for reference and test utterances spoken over the telephone, satisfactory verification is realized only when all the orthogonal parameters are utilized in distance computation. It is finally shown that a significant improvement in accuracy is realizable if the test utterances (whether corrupted by additive noise or by passage through noisy telephone channels) are preprocessed by a modified adaptive noise cancellation filter.