Isolated Letter Recognition Research Articles

Design of a real time automatic speech recognition system using Modified One Against All SVM classifier

In this paper, Texas Instruments TMS320C6713 DSP based real-time speech recognition system using Modified One Against All Support Vector Machine (SVM) classifier is proposed. The major contributions of this paper are: the study and evaluation of the performance of the classifier using three feature extraction techniques and proposal for minimizing the computation time for the classifier. From this study, it is found that the recognition accuracies of 93.33%, 98.67% and 96.67% are achieved for the classifier using Mel Frequency Cepstral Coefficients (MFCC) features, zerocrossing (ZC) and zerocrossing with peak amplitude (ZCPA) features respectively. To reduce the computation time required for the systems, two techniques – one using optimum threshold technique for the SVM classifier and another using linear assembly are proposed. The ZC based system requires the least computation time and the above techniques reduce the execution time by a factor of 6.56 and 5.95 respectively. For the purpose of comparison, the speech recognition system is also implemented using Altera Cyclone II FPGA with Nios II soft processor and custom instructions. Of the two approaches, the DSP approach requires 87.40% less number of clock cycles. Custom design of the recognition system on the FPGA without using the soft-core processor would have resulted in less computational complexity. The proposed classifier is also found to reduce the number of support vectors by a factor of 1.12–3.73 when applied to speaker identification and isolated letter recognition problems. The techniques proposed here can be adapted for various other SVM based pattern recognition systems.

An enhanced dynamic time warping model for improved estimation of DTW parameters

We introduce an enhanced dynamic time warping model (EDTW) which, unlike conventional dynamic time warping (DTW), considers all possible alignment paths for recognition as well as for parameter estimation. The model, for which DTW and the hidden Markov model (HMM) are special cases, is based on a well-defined quality measure. We extend the derivation of the Forward and Viterbi algorithms for HMMs, in order to obtain efficient solutions for the problems of recognition and optimal path alignment in the new proposed model. We then extend the Baum-Welch (1972) estimation algorithm for HMMs and obtain an iterative method for estimating the model parameters of the new model based on the Baum inequality. This estimation method efficiently considers all possible alignment paths between the training data and the current model. A standard segmental K-means estimation algorithm is also derived for EDTW. We compare the performance of the two training algorithms, with various path movement constraints, in two isolated letter recognition tasks. The new estimation algorithm was found to improve performance over segmental K-means in most experiments.

Simulation of artificial vision: I. Eccentric reading of isolated words, and perceptual learning

Simulations of artificial vision were performed to assess “minimum requirements for useful artificial vision”. Retinal prostheses will be implanted at a fixed (and probably eccentric) location of the retina. To mimic this condition on normal observers, we projected stimuli of various sizes and content on a defined stabilised area of the visual field. In experiment 1, we asked subjects to read isolated 4-letter words presented at various degrees of pixelisation and at various eccentricities. Reading performance dropped abruptly when the number of pixels was reduced below a certain threshold. For central reading, a viewing area containing about 300 pixels was necessary for close to perfect reading (>90% correctly read words). At eccentricities beyond 10°, close to perfect reading was never achieved even if more than 300 pixels were used. A control experiment using isolated letter recognition in the same conditions suggested that lower reading performance at high eccentricity was in part due to the “crowding effect”. In experiment 2, we investigated whether the task of eccentric reading under such specific conditions could be improved by training. Two subjects, naive to this task, were trained to read pixelised 4-letter words presented at 15° eccentricity. Reading performance of both subjects increased impressively throughout the experiment. Low initial reading scores (range 6%–23% correct) improved impressively (range 64%–85% correct) after about one month of training (about 1 h/day). Control tests demonstrated that the learning process consisted essentially in an adaptation to use an eccentric area of the retina for reading. These results indicate that functional retinal implants consisting of more than 300 stimulation contacts will be needed. They might successfully restore some reading abilities in blind patients, even if they have to be placed outside the foveal area. Reaching optimal performance may, however, require a significant adaptation process.

Unsupervised speaker adaptation in feature‐based isolated letter recognition

This paper describes two new methods by which a feature‐based isolated letter recognition system, FEATURE [Cole et al., Proc. ICASSP‐1983, pp. 731–733], can learn the acoustical characteristics of individual speakers without feedback from the user. The original tuning algorithm of FEATURE made use of labeled observations from a given speaker, and of the a priori covariances of the feature mean values within the across decision classes, to optimally adjust the expected feature values of all the decision classes. In the first new procedure the system assumes a correct decision every time it classifies a sound with a high confidence level. The class‐conditional mean vectors and covariance matrices are then computed in the same way as in the tuning‐with‐feedback procedure. In the second new procedure the system records the average observed feature values over all possible classes on a feature‐by‐feature basis. The a priori covariances between the mean values of a feature for each decision class with its mean values over all possible classes are then used to update the class‐conditional means and covariances. Experiments were conducted on confusable sets of letters using both speaker adaptation procedures. A significant improvement in letter recognition performance using both analyses was observed after a small number of iterations. [Supported by NSF and DARPA.]

Tuning to the speaker: dynamic adaptation of statistical parameters in isolated letter recognition

The performance of a feature‐based, speaker‐independent recognition system can be improved by enabling the system to learn the acoustical characteristics of individual speakers. Even when features are designed to be speaker‐independent, it is typically observed that for a given feature and pair of letters within‐speaker variation can be less than between‐speaker variation. For example, across all speakers a given feature may have an expected value of 5 for the letter M and 10 for N, but a certain speaker may produce average values of 9 for M and 14 for N. In such cases it is necessary to adjust statistical feature parameters to the individual speaker to obtain optimal recognition performance. This paper describes a dynamic adaptation procedure for updating expected feature values during recognition. The algorithm uses maximum a posteriori probability (MAP) estimation techniques to update the mean vectors of sets of measure values on a speaker‐by‐speaker basis. In updating these mean vectors, the algorithm makes use of the observations input thus far, the relative variability of the features' means within and across subjects, and the covariance of the mean vectors within and across the various letters or sets of letters. The use of tuning produced a dramatic decrease in error rates for certain speakers and letters. [Work supported by NSF.]

Decisions about features

This paper describes the methods of statistical analysis used to classify letters in a feature‐based, speaker‐independent isolated letter recognition system. A hierarchical decision structure was implemented so that decisions at each node of the decision tree could be made on the basis of a small number of relevant features. For example, the 26 letters were first classified into vowel categories on the basis of first and second formant frequencies. The specific decisions, and the features used to make them, were selected by a clustering analysis of training data. At each decision node of the recognition system the test utterance was first analyzed using Fisher linear discriminant functions, with threshold weights individually set for each pairwise decision in order to minimize misclassifications. When a decision could not be made with certainty, classification was performed using a maximum likelihood procedure assuming multivariate Gaussian statistics. The sequential use of nonparametric and parametric discriminant functions produced superior classification to that obtained with either of the separate analyses. The overall system structure will be discussed in terms of practical tradeoffs between the number of features used at each decision node and the system's overall probability of error. [Work supported by NSF.]