An exploratory implementation of a syllable-based recognizer is described. Continuous speech is first divided into syllabic unite, and the units are then matched against syllable templates using a dynamic programming algorithm. A hierarchical transition network is used to limit the syllable search to possible continuations of the current partial sentence hypotheses. Competing hypotheses are pruned by a 'beam search'.Experiments are reported on automatic recognition of English sentences with a 70-word vocabulary and restricted syntax produced by one male speaker. 85% of the sentences were correctly recognized.Comparable results were obtained for a similar task in French using a female speaker. The method is computationally efficient: real-time performance on dedicated hardware should bs obtainable without difficulty. A method of scaling the distance measures used in the syllable matching is described. This scaling takes into account variability in syllable production, both as a function of position within the syllable and as a function of the various spectral parameters being used. 1, INTRODUCTIONAt some level in a continuous speech recognition system there is an interface between acoustically derived information about the speech to be recognized and syntactic constraints on the interpretation of that information. For maximum efficiency in the application of the constraints, the interface level should be as low as possible. At the lowest levels, however, the interpretation of acoustic information is highly context dependent. The syllable level is attractive because it is low enough to make the inventory of units to be recognized reasonably small while being high enough to make the phonetic interpretation of each unit reasonably context independent.In addition, syllable boundaries cam be determined independently of the task of identifying the syllable. This property, which does not hold for words or for phonemes, makes for efficient comparison of syllables with stored reference forms and for efficient evaluation of competing sentence hypotheses. The theoreticsl advantages of the syllable-level representation are discussed by Msrmslstein (1) and Fujimura (2). Davis and Mermelstein (3) report results on syllable-based recognition assisted by lexical information but using CH1559-418010000-0880$O0.75 © 1980 IEEE 880 hand-segmented speech data. Zwicker, Terhardt and Psulus (4) use a. similar syllable-based psychoacoustics-oriented scheme for the recognition of German city names.The experiments reported pertain to speakerdependsnt recognition of pra-specified sentences produced with a limited vocabulary and restricted syntax. One male speaker was used to train and test a 70-word vocabulary English system intended to recognize 'date and time' sentences, e.g., "January fourth at eight forty-five p.m.". A female speaker was used to test and train a conpsrabls French systen. The French vocabulary comprised 43 words and an additional 25 word variants corresponding to sequences of two or more words in which syllables had been ...
This paper presents a new class of A* algorithms for Viterbi phonetic decoding subject to lexical constraints. This type of algorithm can be made to run substantially faster than the Viterbi algorithm in an isolated word recognizer having a vocabulary of 1600 words. In addition, multiple recognition hypotheses can be generated on demand and the search can be constrained to respect conditions on phone durations in such a way that computational requirements are substantially reduced.Results are presented on a 60 OOO word recognition task.
In this study we demonstrate the effectiveness of phonemic hidden Markov models with Gaussian mixture output densities (mixture HMM's) for speaker-dependent large-vocabulary word recognition. Speech recognition experiments show that for almost any reasonable amount of training data, recognizers using mixture HMM's consistently outperform those employing unimodal Gaussian HMM's. With a sufficiently large training set (e.g., more then 2500 words), use of HMM's with 25-component mixture distributions typically reduces recognition errors by about 40%. We also found that the mixture HMM's outperform a set of unimodal generalized triphone models having the same number of parameters. Previous attempts to employ mixture HMM's for speech recognition proved discouraging because of the high complexity and computational cost in implementing the Baum-Welch training algorithm. We show how mixture HMM's can be implemented very simply in the unimodal transition-based frameworks by the device of allowing multiple transitions from one state to another.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.