Speaker verification technology has been successfully adopted and integrated into many applications. However, most of these applications require a microphone located near the talker. For the case of distant microphones, speech signals are corrupted by reverberations caused by the large speaker to microphone distance. In this paper, we first introduce a new feature set that gives more details in the frequency dimension in the 2-D time-frequency space used to represent speech. These features are computed using two sets of basis vectors, both of which are applied directly to the amplitude compressed FFT spectrum. One set of basis vectors accounts for the spectral envelope while the second set accounts for pitch. Those features are used to train a Convolutional Neural Network (CNN), with the goal of reducing the negative effects of reverberation. The proposed frontend is shown to be robust for speaker verification in reverberant environments.

In this paper, a feature set referred to as Discrete Cosine Series (DCS) is proposed for noise robust Automatic Speech Recognition (ASR). Unlike many other robust algorithms which use various forms of “long term” processing, DCS uses a small frame spacing to facilitate separating speech from noise and also for other benefits. Spectral and temporal modulations are performed separately using only a small number of modulation filters. ASR experiments show the effectiveness of individual components of the DCS algorithm. The DCS features yield higher accuracy ASR for both additive noise and reverberation, as compared to several other advanced robust algorithms.

For many years, filterbank have been widely used as one step of frontend feature extraction for Automatic Speech Recognition (ASR). In this paper, we propose a unified framework for ASR frontends, by first moving the nonlinear amplitude scaling, and then combining the filterbank weights with the cosine basis vectors. As part of this framework, we also show that the delta terms used to encode feature dynamics can also be viewed as one realization of a set of “unified” basis vectors over time. With this framework, frontends can be developed, analyzed and evaluated through their basis vectors over frequency and time.

In this paper, the ability of human listeners to recognize tones from continuous Mandarin Chinese is evaluated and compared to the accuracy of automatic systems for tone classification and recognition. All tones used for experimentation were extracted from the RASC863 continuous Mandarin Chinese database. The human listeners are native speakers of Mandarin and the automatic methods consist of tone classification using neural networks and tone recognition using Hidden Markov Models. Features used for the automatic methods are a combination of spectral/temporal features, energy contours, and pitch contours. When very little context is used (i.e., vowel segments only) the human and machine performance is comparable. However, as the context interval is increased, the human performance is much better than the best machine performance obtained.

Many speech segmentation techniques have been proposed to automate phonetic alignment. Most of the techniques require, however, labeled data to train, and perform well only for read, high-quality speech. Automatic phonetic alignment, for lower quality varied data with no labeled training data, the subject of this paper, is a much more challenging domain. An HMMbased automatic speech recognizer was used in this study to determine phonetic sequences and boundaries of “open source” speech data, retrieved from public websites. The HMM models were initially trained using the TIMIT database and subsequently adapted to each passage. Standard frontend features such as MFCC, LPCC and PLP, and features computed by applying the DCT directly to the short-time spectrum (DCTC) were evaluated using TIMIT data. The “best” parameter set was found to be DCTC_78 and these parameters were used to align the speech data of interest.

This paper presents two nonlinear feature dimensionality reduction methods based on neural networks for a HMM-based phone recognition system. The neural networks are trained as feature classifiers to reduce feature dimensionality as well as maximize discrimination among speech features. The outputs of different network layers are used for obtaining transformed features. Moreover, the training of the neural networks uses the category information that corresponds to a state in HMMs so that the trained networks can better accommodate the temporal variability of features and obtain more discriminative features in a low dimensional space. Experimental evaluation using the TIMIT database shows that recognition accuracies with the transformed features are slightly higher than those obtained with original features and considerably higher than obtained with linear dimensionality reduction methods. The highest phone accuracy obtained with 39 phone classes and TIMIT was 74.9% using a large number of training iterations based on the state-specific targets.

Recently, the modulation spectrum has been proposed and found to be a useful source of speech information. The modulation spectrum represents longer term variations in the spectrum and thus implicitly requires features extracted from much longer speech segments compared to MFCCs and their delta terms. In this paper, a Discrete Cosine Transform (DCT) analysis of the log magnitude spectrum combined with a Discrete Cosine Series (DCS) expansion of DCT coefficients over time is proposed as a method for capturing both the spectral and modulation information. These DCT/DCS features can be computed so as to emphasize frequency resolution or time resolution or a combination of the two factors. Several variations of the DCT/DCS features were evaluated with phonetic recognition experiments using TIMIT and its telephone version (NTIMIT). Best results obtained with a combined feature set are 73.85 % for TIMIT and 62.5 % for NTIMIT. The modulation features are shown to be far more important than the spectral features for automatic speech recognition and far more noise robust. Index Terms: modulation spectrum, spectral-temporal feature,

A neural network based feature dimensionality reduction for speech recognition is described for accurate phonetic speech recognition. In our previous work, a neural network based nonlinear principal component analysis (NLPCA) was proposed as a dimensionality reduction approach for speech features. It was shown that the reduced dimensionality features are very effective for representing data for vowel classification. In this paper, we extend this neural network based NLPCA approach for phonetic recognition using continuous speech. The reduced dimensionality features obtained with NLPCA are used as the features for HMM phone models. Experimental evaluation using the TIMIT database shows that recognition accuracies with NLPCA reduced dimensionality features are higher than recognition rates obtained with original features, especially when a small number of states and mixtures are used for HMM phonetic models. Index Terms: feature dimensionality reduction, neural networks, HMMs, speech recognition

One of the main practical difficulties for automatic speech recognition is the large dimensionality of acoustic feature spaces and the subsequent training problems collectively referred to as the “curse of dimensionality. ” Many linear techniques, most notably principal components analysis (PCA) and linear discriminant analysis (LDA) and several variants have been used to reduce dimensionality while attempting to preserve variability and discriminability of classes in the feature space. However, these orthogonal rotations of the feature space are suboptimal if data are distributed primarily on curved subspaces embedded in the higher dimensional feature spaces. In this paper, two neural network based nonlinear transformations are used to represent speech data in reduced dimensionality subspaces. It is shown that a subspace computed with the explicit intent of maximizing classification accuracy is far superior to a subspace derived as to minimize mean square representation error. Index Terms: dimensionality reduction, nonlinear principal components analysis

This paper presents speech signal modeling techniques that are well suited to robust recognition of connected digits in noisy environments. After several preprocessing steps speech is represented by a block-encoding of discrete cosine transform of its spectra. In this paper we combine linear predictive coding (LPC), morphological filtering, and long block lengths to achieve robust features for improved recognition in noisy environments. The spectral envelope is first estimated by LPC. Subsequent morphological filtering enhances the peaks while smoothing the valleys, which are more affected by noise in the signal. These techniques were tested with the Aurora 2 database and the standard HMM recognizer as defined by the ETSI STQ-AURORA DSR Working group for WI007. With no major increase in computational demand a 23% word error rate (WER) reduction has been achieved as compared to the WI007 baseline MFCC front-end for multi-condition training condition. The basic conclusion is that the features resulting from the methods presented here perform better than cepstral features for ASR of noisy speech.

In this paper, a new spectral/temporal method is described for robust pitch tracking for both high quality and telephone speech. A previous version of this algorithm was presented as YAAPT (Kasi and Zahorian, 2002) [10]. In the current paper, a novel method is presented for spectral pitch tracking, using nonlinear processing to partially restore the potentially missing fundamental frequency. A frequency domain modified autocorrelation is used to determine the spacing between harmonic peaks in the spectrum. The frequency domain spectral track is then used to refine time-domain pitch candidates obtained using the “NCCF or Normalized Cross Correlation” reported by Talkin [1]. Dynamic programming is used to find the “best” pitch track among all the candidates, using both local and transition costs. The algorithm was evaluated using the Keele pitch extraction reference database.

The paper describes an algorithm for determining pitch period markers in a continuous speech signal using prior knowledge of pitch values. The algorithm uses dynamic programming to determine optimal markers from a set of probable markers. Local costs are assigned based on the amplitudes of local peaks, and transition costs are based on the closeness of peak spacings to the spacing predicted by the pitch period. In tests conducted with the Keele database, it was found that over 90% of pitch markers are within /spl plusmn/0.3 ms of reference marks in the laryngograph signal recorded with the Keele acoustic data, and over 96% of markers are within /spl plusmn/1 ms of reference markers. One of the features of this algorithm is tolerance to errors in the pitch track. The algorithm can tolerate errors in the pitch track of approximately -10% to +60%, with very little degradation in performance. It also generates 1% extraneous peaks.

Automatic speech recognizers perform poorly when training and test data are systematically different in terms of noise and channel characteristics. One manifestation of such differences is variations in the probability density functions (pdfs) between training and test features. Consequently, both automatic speech recognition and automatic speaker identification may be severely degraded. Previous attempts to minimize this problem include Cepstral Mean and Variance Normalization and transforming all speech features to a univariate Gaussian pdf. In this paper, we present a quantile based Cumulative Density Function (CDF) matching technique for data drawn from different distributions. This method can be used to compensate for the systematic marginal (i.e. each feature individually) differences between training and test features. We further propose a linear covariance normalization technique to compensate for differences in covariance properties between training and test data. Experimental results are given that illustrate these techniques for speech recognition and automatic speaker identification.

Two techniques for adaptation based on frequency scale modifications are described and evaluated. In one method, minimum mean square error matching is performed between a spectral template for each to a typical speaker spectral template. One parameter, a warping factor, is used to control the spectral matching. In the second method, a neural network classifier is used to adjust the frequency warping factor for each so as to maximize vowel classification performance for each speaker. A vowel classifier trained only with normalized female speech and tested only with normalized male speech, or vice versa, is nearly as accurate as when genders are matched for training and testing, and the speech is not normalized. The improvement due to normalization is much smaller, if training and test data are matched. The normalization based on classification performance is superior to that based on minimizing mean square error.

A visual speech training aid for persons with hearing impairments has been developed using a Windows-based multimedia computer. Previous papers (Zahorian, S. et al., Int. Conf. on Spoken Language Processing, 2002; Zahorian and Nossair, Z.B., IEEE Trans. on Speech and Audio Processing, vol.7, no.4, p.414-25, 1999; Zimmer, A. et al., ICASSP, vol.6, p.3625-8, 1998; Zahorian and Jagharghi, A., J. Acoust. Soc. Amer., vol.94, no.4, p.1966-82, 1993) have describe the signal processing steps and display options for giving real-time feedback about the quality of pronunciation for 10 steady-state American English monopthong vowels (/aa/, /iy/, /uw/, /ae/, /er/, /ih/, /eh/, /ao/, /ah/, and /uh/). This vowel training aid is thus referred to as a vowel articulation training aid (VATA). We now describe methods to develop a monophone-based hidden Markov model/neural network recognizer such that real time visual feedback can be given about the quality of pronunciation of short words and phrases. Experimental results are reported which indicate a high degree of accuracy for labeling and segmenting the CVC (consonant-vowel-consonant) database developed for "training" the display.

A visual speech training aids for persons with hearing impairments has been developed using a Windows-based multimedia computer. In previous papers, the signal processing steps and display options have been described for giving real-time feedback about the quality of pronunciation for 10 steady-state American English monopthong vowels (/aa/, /iy/, /uw/, /ae/, /er/, /ih/, /eh/, /ao/, /ah/, and /uh/). This vowel training aid is thus referred to as a vowel articulation training aid (VATA). In the present paper, methods are described to develop a monophone-based hidden Markov model/neural network recognizer such the real-time visual feedback can be given about the quality of pronunciation of short words and phrases. Experimental results are reported which indicate a high degree of accuracy for labeling and segmenting the CVC database developed for "training" the display.

Teager Energy and Energy-Entropy Features are two approaches, which have recently been used for locating the endpoints of an utterance. However, each of them has some drawbacks for speech in noisy environments. This paper proposes a novel method to combine these two approaches to locate endpoint intervals and yet make a final decision based on energy, which requires far less time than the feature based methods. After the algorithm description, an experimental evaluation is presented, comparing the automatically determined endpoints with those determined by skilled personnel. It is shown that the accuracy of this algorithm is quite satisfactory and acceptable.

In this paper, we present a pitch detection algorithm that is extremely robust for both high quality and telephone speech. The kernel method for this algorithm is the “NCCF or Normalized Cross Correlation” reported by David Talkin [1]. Major innovations include: processing of the original acoustic signal and a nonlinearly processed version of the signal to partially restore very weak F0 components; intelligent peak picking to select multiple F0 candidates and assign merit factors; and, incorporation of highly rohust pitch contours obtained from smoothed versions of low frequency portions of spectrograms. Dynamic programming is used to find the “best” pitch track among all the candidates, using both local and transition costs. We evaluated our algorithm using the Keele pitch extraction reference database as “ground truth” for both “high quality” and “telephone” speech. For both types of speech, the error rates obtained are lower than the lowest reported in the literature.

A visual speech training aid for persons with hearing impairments has been developed using a Windows-based multimedia computer. The training aid provides real time visual feedback as to the quality of pronunciation for 10 steady-state American English monopthong vowels (/aa/, /iy/, /uw/, /ae/, /er/, /ih/, /eh/, /ao/, /ah/, and /uh/). This training aid is thus referred to as a Vowel Articulation Training Aid (VATA). Neural network (NN) classifiers are used to classify vowels and then provide real time feedback for several displays: a 10-category “vowel bargraph ” which provides “discrete ” feedback, an “ellipse display ” which provides continuous feedback over a 2-D space similar to a formant1-formant2 space, and three game displays (a form of "tetrus," controlled by one vowel, a "chicken crossing the road," controlled by two vowels, and pacman, controlled by four vowels.) Continuous feedback such as this is desirable for speech training to help improve articulation. In this paper we describe the overall speech training system, discuss some algorithmic refinements to the vowel classifier, and report some experiments related to the development of a database used for "training " the display.

This paper describes speech signal modeling techniques which are well-suited to high performance and robust isolated word recognition. We present new techniques for incorporating spectral/temporal information as a function of the temporal position within each word. In particular, spectral/temporal parameters are computed using both variable length blocks with a variable spacing between blocks. We tested features computed with these methods using an alphabet recognition task based on the ISOLET database. The hidden Markov model toolkit (HTK) was used to implement the isolated word recognizer with whole word HMM models. The best accuracy achieved for speaker independent alphabet recognition, using 50 features, was 97.9%, which represents a new benchmark for this task. We also tested these methods with deliberate signal degradation using additive Gaussian noise and telephone band limiting and found that the recognition degrades gracefully and to a smaller degree than for control cases based on MFCC coefficients and delta cepstra terms.

This paper presents an investigation of non-uniform time sampling methods for spectral/temporal feature extraction for use in automatic speech recognition. In most current methods for signal modeling of speech information, “dynamic ” features are determined from frame-based parameters using a fixed time sampling, i.e., fixed block length and fixed block spacing. This work explores new methods in which block length and/or block spacing are variable. Three methods are suggested and each was tested with the TIMIT database using a standard HMM recognizer. Phone recognition experiments were conducted using the standard 39 phone set. The methods were also evaluated with various HMM model complexities. Experimental results indicated that none of the proposed nonuniform feature time sampling methods perform significantly better than fixed time sampling methods. However, the best results obtained with the front end are comparable to those obtained with current state-of-the-art systems. Also the performance of our monophone system surpasses that of most reported context-dependent monophone systems. 1.

Spectral feature computations continue to be a very difficult problem for accurate machine recognition of vowels especially in the presence of noise or for otherwise degraded acoustic signals. In this work, a new peak envelope method for vowel classification is developed, based on a missing frequency components model of speech recognition. According to this model, vowel recognition depends only on the location of spectral peaks. Also, smoothing and interpolation of the sampled spectra, performed in the cepstral analysis method commonly used in automatic speech recognition results in a loss of valuable information. The new method for feature extraction presented in this paper is based on minimum mean square error curve fitting of cosine-like basis vectors to all peaks in the speech spectrum. A mathematical model for smoothly tracking spectral envelopes using only spectral peak information and ignoring other parts of the spectrum is presented. A software algorithm for the model was developed and tested for various speaker types using a neural network classifier. Vowel classification experiments were conducted based on the features derived from the spectral peaks. The classification rates of the peak method under various signal to noise ratios was also evaluated. The basic conclusion is that the new features perform the same as cepstral features for clean speech, but have advantages when the signal is degraded by noise.

This paper presents speech signal modeling techniques which are well suited to high performance and robust isolated word recognition. Speech is encoded by a discrete cosine transform of its spectra, after several preprocessing steps. Temporal information is then also explicitly encoded into the feature set. We present a new technique for incorporating this temporal information as a function of the temporal position within each word. We tested features computed with this method using an alphabet recognition task based on the ISOLET database. The HTK toolkit was used to implement the isolated word recognizer with whole word HMM models. The best result obtained based on 50 features and speaker independent alphabet recognition was 98.0%. Gaussian noise was added to the original speech to simulate a noisy environment. We achieved a recognition accuracy of 95.8% at a SNR of 15 dB. We also tested our recognizer with simulated telephone quality speech by adding noise and band limiting the original speech. For this "telephone" speech, our recognizer achieved 89.6% recognition accuracy. The recognizer was also tested in a speaker dependent mode, resulting in a 97.4% accuracy on test data.

A neural network algorithm for speaker identification with large groups of speakers is described. This technique is derived from a technique in which an N-way speaker identification task is partitioned into N*(N-1)/2 two-way classification tasks. Each two-way classification task is performed using a small neural network which is a two-way, or pair-wise, network. The decisions of these two-way networks are then combined to make the N-way speaker identification decision (Rudasi and Zahorian, 1991 and 1992). Although very accurate, this method has the drawback of requiring a very large number of pair-wise networks. In the new approach, two-way neural network classifiers, each of which is trained only to separate two speakers, are also used to separate other pairs of speakers. This method is able to greatly reduce the number of pair-wise classifiers required for making an N-way classification decision, especially when the number of speakers is very large. For 100 speakers extracted from the TIMIT database, the number of pair-wise classifiers can be reduced by approximately a factor of 5, with only minor degradation in performance when 3 seconds or more of speech is used for identification. Using all 630 speakers from the TIMIT database, this method can be used to obtain over 99.7% accuracy. With the telephone version of the same database, an accuracy of 40.2% can be obtained.

A novel pattern classification technique and a new feature extraction method are described and tested for vowel classification. The pattern classification technique partitions an N-way classification task into N*(N-1)/2 two-way classification tasks. Each two-way classification task is performed using a neural network classifier that is trained to discriminate the two members of one pair of categories. Multiple two way classification decisions are then combined to form an N-way decision. Some of the advantages of the new classification approach include the partitioning of the task allowing independent feature and classifier optimization for each pair of categories, lowered sensitivity of classification performance on network parameters, a reduction in the amount of training data required, and potential for superior performance relative to a single large network. The features described in this paper, closely related to the cepstral coefficients and delta cepstra commonly used in speech analysis, are developed using a unified mathematical framework which allows arbitrary nonlinear frequency, amplitude, and time scales to compactly represent the spectral/temporal characteristics of speech. This classification approach, combined with a feature ranking algorithm which selected the 35 most discriminative spectral/temporal features for each vowel pair, resulted in 71.5% accuracy for classification of 16 vowels extracted from the TIMIT database. These results, significantly higher than other published results for the same task, illustrate the potential for the methods presented in this paper.

A vowel training aid system for hearing impaired persons which uses a Windows-based multimedia computer has been developed. The system provides two main displays which give visual feedback for vowels spoken in isolation and short word contexts. Feature extraction methods and neural network processing techniques provide a high degree of accuracy for speaker independent vowel training. The system typically provides correct classification of over 85% of steady state vowels spoken by adult male, adult female and child (both genders combined) speakers. Similar classification accuracy is also observed for vowels spoken in short words. Low cost and good performance make this system potentially useful for speech training at home.

Spectral/temporal segment features are adapted for isolated word recognition and tested with the entire English alphabet set using Hidden Markov Models.The ISOLET database from OGI and the HTK toolkit from Cambridge university were used to test our feature extraction technique.With our feature set we were able to achieve 97.3% recognition accuracy on test data with one pass using a whole word based recognizer.Gaussian noise was also added to evaluate robustness of the feature set.We were able to obtain recognition accuracies of 49.6% and 84.3% at SNR of -10dB and 0dB, respectively.Linear discriminant analysis was also applied to the initial feature set for a number of feature configurations and noise levels but, generally, the performance was not improved.We conclude that the initial feature computations used are both very efficient (best results obtained with 50 total features) and robust in the presence of noise.

This paper presents methods and experimental results for phonetic classification using 39 phone classes and the NIST recommended training and test sets for NTIMIT and TIMIT. Spectral/temporal features which represent the smoothed trajectory of FFT derived speech spectra over 300 ms intervals are used for the analysis. Classification tests are made with both a binary-pair partitioned (BPP) neural network system (one neural network for each of the 741 pairs of phones) and a single large neural network. The classification accuracy is very similar for the two types of networks, but the BPP method has the advantage of a much shorter training time. The best results obtained (77% for TIMIT and 67.4% for NTIMIT) compare favorably to the best results reported in the literature for this task.

Experiments in modeling speech signals for phoneme classification are described. Enhancements to standard speech processing methods include basis vector representations of dynamic feature trajectories, morphological smoothing (dilation) of spectral features, and the use of many closely spaced, short analysis windows. Results are reported from experiments using the TIMIT database of up to 71.0% correct classification of 16 presegmented vowels in a noise-free environment, and 54.5% correct classification in a 10 dB signal-to-noise ratio environment.

A spoken language system combines speech recognition, natural language processing and human interface technology. It functions by recognizing the person's words, interpreting the sequence of words to obtain a meaning in terms of the application, and providing an appropriate response back to the user. Potential applications of spoken language systems range from simple tasks, such as retrieving information from an existing database (traffic reports, airline schedules), to interactive problem solving tasks involving complex planning and reasoning (travel planning, traffic routing), to support for multilingual interactions. We examine eight key areas in which basic research is needed to produce spoken language systems: (1) robust speech recognition; (2) automatic training and adaptation; (3) spontaneous speech; (4) dialogue models; (5) natural language response generation; (6) speech synthesis and speech generation; (7) multilingual systems; and (8) interactive multimodal systems. In each area, we identify key research challenges, the infrastructure needed to support research, and the expected benefits. We conclude by reviewing the need for multidisciplinary research, for development of shared corpora and related resources, for computational support and far rapid communication among researchers. The successful development of this technology will increase accessibility of computers to a wide range of users, will facilitate multinational communication and trade, and will create new research specialties and jobs in this rapidly expanding area.>

A vowel articulation training aid for the deaf has been developed. This is an improved digital system processor (DSP) version of the analog filter-bank system presented by S.A. Zahorian and S. Vankat (1990). The real-time signal processing strategy and the mapping of the continuous feature space to a two-dimensional display space are described. A combined linear/nonlinear transformation, in the form of an artificial neural network, is used to convert cepstral coefficients to a 2-D space suitable for computer displays. A multistage artificial neural network training algorithm has been developed that yields better performance and reduced training time relative to conventional training techniques.>

The authors introduce a novel method for partitioning a large classification problem using N*(N-1)/2 binary pair classifiers. The binary pair classifier has been applied to a speaker identification problem using neural networks for the binary classifiers. The binary partitioned approach was used to develop an identification system for the 47 male speakers belonging to the Northern dialect region of the TIMIT database. The system performs with 100% accuracy in a text-independent mode when trained with about 9 to 14 s of speech and tested with 8 s of speech. The partitioned approach performs comparably, or even better, than a single large neural network. For large values of N (>10), the partitioned approach requires only a fraction of the training time required for a single large network. For N=47, the training time for the partitioned network would be about two orders of magnitude less than for the single large network.>

The authors present a method for improving HMM (hidden Markov model) phonetic discrimination capability which is a linear discriminant transform of acoustic features to a continuous-valued feature space such that phonetic distinctions correlate closely with Euclidean distance in the transformed feature space. Experimental testing with a 30-word single syllable highly confusable vocabulary showed that the acoustic-phonetic transform could be used to reduce word error rates approximately 25%. In general, results based on the LDA2 transform, i.e., linear discriminant analysis with whitening of the within-class covariance matrices, are superior to those obtained with LDA1, linear discriminant analysis without whitening. Recognition results also improve if a block transform of several frames per block is used rather than a transform based on one frame per block.>

Several acoustic feature sets and automatic classifiers were investigated to determine a combination of features and classifiers which would permit accurate bottom-up speaker- and vowel-independent automatic recognition of initial stop consonants in English. The features evaluated included a form of cepstral coefficients and formants, each computed both for one static frame and as spectral trajectories over various segments of the speech signal. The classifiers investigated included Bayesian maximum-likelihood (BML), artificial neural network (NN), and K-nearest-neighbor (KNN) classifiers. The most accurate results, over 93% of the six stops correctly identified with a speaker-independent classifier, were obtained with the BML classifier using cepstral coefficient trajectories as a 20-dimensional feature vector. These results for stop recognition are higher than any results previously reported for a database of similar diversity.>

The signal processing strategy for a computer-based vowel articulation training aid for the deaf is described. This processing is based on a nonlinear/linear artificial neural network transformation of 16-channel filter bank data to a two-dimensional space which approximates an F1/F2 space. Speaker-independent vowel training displays have been developed with vowel identity cued by spatial location and color. Testing with both normally-hearing and hearing-impared listeners indicates that the display is very easy to interpret and that the relationship between the pattern and the spoken vowel is consistent. The continuous relationship between phonetic perception and display patterns provides feedback for fine tuning of vocal tract settings.>

Speech was analyzed using a cosine basis vector set to model speech spectral magnitude. This basis vector set is quite similar to an optimal principal-components basis vector set. However, in contrast with the principal components, under some conditions the cosine basis vector coefficients can be used to directly determine an FIR speech synthesis filter. Depending on the details of how the magnitude spectrum is first scaled in frequency and in amplitude before the cosine basis vector modeling, different forms of FIR filters are obtained. Speech was synthesized using the FIR filters obtained from discrete cosine coefficients and compared with speech synthesized from a principal-components-based vocoder, and with speech synthesized from a linear predictive vocoder. For a fixed number of synthesis filter parameters, the FIR synthesis with a minimum phase characteristic was nearly as good as synthesis based on the optimal principal components and about the same as linear predictive synthesis.