Phoneme Recognition Rate Research Articles

A recurrent error propagation network speech recognition system

This paper describes a speaker-independent phoneme and word recognition system based on a recurrent error propagation network (REPN) trained on the TIMIT database. The REPN is a fully recurrent error propagation network trained by the propagation of the gradient signal backwards in time. A variation of the stochastic gradient descent procedure is used which updates the weights by an adaptive step size in the direction given by the sign of the gradient. Phonetic context is stored internal to the network and the outputs are estimates of the probability that a given frame is part of a segment labelled with a context-independent phonetic symbol. During recognition, a dynamic programming match is made to find the most probable string of symbols. The one pass algorithm is used for phoneme and word recognition. The phoneme recognition rate for all 61 TIMIT symbols is 70·0% correct (63·5% accuracy including insertion errors) and on a reduced 39-symbol set the recognition rate is 76·5% correct (69·8%). This compares favourably with the results of other methods, such as HMMs, on the same database [K. F. Lee & H. W. Hon 1989. IEEE Transactions on Acoustics, Speech and Signal Processing, 37, 1641–1648; S. E. Levinson, M. Y. Liberman, A. Ljolje & L. G. Miller 1989. Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing. Glasgow, pp. 441–444]. Analysis of the phoneme recognition results shows that information available from bigram and durational constraints is adequately handled within the network allowing for efficient parsing of the network output. For comparison, there is less computation involved in the resulting scheme than in a one-state-per-phoneme HMM system. This is demonstrated by applying the recognizer to the DARPA 1000-word resource management task. Parsing the network output to the word level with no grammar and no pruning can be carried out in faster than real time on a SUN 4 330 workstation.

Phoneme recognition with elliptic discrimination neural units

AbstractMany researchers achieved high phoneme recognition rates by multilayered neural networks with linear discrimination neural (LDN) units. However, it is difficult to analyze which components of the input are important to each unit in those LDN networks.This paper proposed a multilayer neural network with elliptic discrimination neural (EDN) units so that the functions of each unit in the network may be interpreted more definitely. The center of the elliptic discrimination boundary of a neural unit corresponds to a typical point in an input space. The radii of the ellipse express the extent of the corresponding components in the input space, hence it becomes clear which components of the input space are important to each unit in the EDN network.To compare the performance of EDN and LDN networks, recognition experiments of phonemes /b, d, g/ in 5240 tokens of a Japanese speech database were carried out. In the experiments, recognition rates were obtained by EDN networks as high as the rate by an LDN network. Also, it was confirmed which components of the input space are important to each unit in the EDN network.

Speech recognition using HMMs with an LVQ-trained codebook

A new speech recognition algorithm using HMMs with a codebook trained by learning vector quantization (LVQ) is here described. Both HMMs and LVQ are stochastic algorithms holding considerable promise for speech recognition. In particular, HMMs have the significant advantage that phone models can easily be concatenated to produce a long utterance model, such as a word or sentence model. LVQ, on the other hand, is a powerful classifier, as shown in the high phoneme recognition rates obtained in McDermott and Katagiri (1989). The new algorithm described here combines the advantages inherent in each of these two algorithms. To evaluate this LVQ-HMM hybrid, phoneme recognition experiments were performed using the same data as used in TDNN (Waibel, 1988) and shift-tolerant LVQ (McDermott and Katagiri, 1989) experiments. Applied to various phoneme recognition tasks, the LVQ-HMM hybrid achieved recognition rates much higher than those of a conventional HMM with a codebook designed by K-means clustering. For small codebook sizes, the hybrid model realized a more than tenfold decrease in recognition error rates, yielding extremely high phoneme recognition performance, comparable to that of TDNN or shift-tolerant LVQ.

Duration control methods for HMM phoneme recognition

Two kinds of duration control for HMM (hidden Markov model) phoneme recognition are proposed: phoneme duration control for an HMM phone model and an event duration control for an HMM state. The phoneme duration control is carried out by combining an HMM output probability with a phoneme duration probability. The phoneme duration probability is calculated using a phoneme duration histogram obtained from training samples. Phoneme duration control is effective in discriminating phonemes with different durations such as /n/ and /N/. Event duration control is realized as a state duration penalty calculated from an HMM state duration probability of training samples. Event duration control is effective in discriminating phonemes with different event structures such as /s/ and /ts/. Recognition experiments are carried out using Japanese phonemes extracted from an isolated word database uttered by one male speaker. The phoneme recognition rate is improved from 84.8%–90.0% using these duration control techniques.

Recognition of Spoken Words and Phrases in Multitalker Environment Using Syntactic Methods

We describe a method of recognizing isolated words and phrases from a given vocabulary spoken by any member in a given group of speakers, the identity of the speaker being unknown to the system. The word utterance is divided into 20-30 nearly equal frames, frame boundaries being aligned with glottal pulses for voiced speech. A constant number of pitch periods are included in each frame. Statistical decision rules are used to determine the phoneme in each frame. Using the string of phonemes from all the frames of the utterance, a word decision is obtained using (phonological) syntactic rules. The syntactic rules used here are of 2 types, namely, 1) those obtained from the theory of word construction from phonemes in English as applied to our vocabulary, 2) those used to correct possible errors in phonemic decisions obtained earlier based on the decisions of neighboring segments. In our experiment, the vocabulary had 40 words, consisting of many pairs of words which are phonemically close to each other. The number of speakers was 6. The identity of the speaker is not known to the system. In testing 400 words utterances, the recognition rate was about 80 percent for phonemes (for 11 phonemes) but the word recognition was 98.1 percent correct. Phonological-syntactic rules played an important role in upgrading the word recognition rate over the phoneme recognition rate.