In this paper, “continuous speech recognition” problem is given a clear mathematical formulation as the search for that sequence of basic speech units that best fits the input acoustic pattern. For this purpose spoken language models in the form of hierarchical transition networks are introduced, where lower level subnetworks describe the basic units as possible sequences of spectral states. The units adopted in this paper are either whole words or smaller subword elements, called diphones. The recognition problem thus becomes that of finding the best path through the network, a task carried out by the linguistic decoder. By using this approach, knowledge sources at different levels are strongly integrated. In this way, early decision making based on partial information (in particular any segmentation operation or the speech/silence distinction) is avoided; usually this is a significant source or errors. Instead, decisions are deferred to the linguistic decoder, which possesses all the necessary pieces of information. The properties that a linguistic decoder must possess in order to operate in real-time are listed, and then a best-few algorithm with partial traceback of explored paths, satisfying the above requisites, is described. In particular, the amount of storage needed is almost constant for any sentence length, the computation is approximately linear with sentence length, and the interpretation of early words in a sentence may be possible long before the speaker has finished talking. Experimental results with two systems, one with words and the other with diphones as basic speech units, are reported. Finally, relative merits of words and diphones are discussed, taking into account aspects such as the storage and computing time requirements, their relative ability to deal with phonological variations and to discriminate between similar words, their speaker adaptation capability, and the ease with which it is possible to change the vocabulary and the language dependencies.

Language models and search algorithms for real-time speech recognition

Computer Models of Speech Using Fuzzy Algorithms

An expert system for mapping acoustic cues into phonetic features

Several methods for acoustic-phonetic decoding are reviewed. Emphasis is placed on the need for mathematical methods for speech recognition. Several examples of statistical methods are described. The author presents several techniques for incorporating “speech knowledge” into these statistical models, and provides a simple formalism for using multiple knowledge sources in a coherent speech recognition system.

Acoustic-phonetic decoding of speech

Disclosed is a behavioral model for wide-band radio frequency transmitters. Also disclosed are various implementation of the behavioral model for purpose of baseband predistortion of dynamic nonlinear systems, such as wideband wireless transmitters and power amplifiers. In one example embodiment, a LBG behavioral model comprises two non-linear cascading modules: a dynamic weak nonlinear (DWNL) module, which models dynamic week nonlinearities of the system, and a static strong nonlinear (SSNL) module, which models static strong nonlinearities of the system. In one example embodiment, a forward LBG model includes DWNL module followed by the SSNL module. In another example embodiment, a reverse LBG model includes SSNL module followed by DWNL module.

Nonlinear behavior models and methods for use thereof in wireless radio systems

The grammars are proposed which generate all possible reference signals of speech using a finite number of elementary reference signals and a finite system of rules. These grammars take into account the essential factors of variability of speech signals: coarticulation of sounds, nonlinear change of the rate and the intensity of pronouncing. The recognition of a shown speech signal consists in finding among all reference signals one which has the greatest resemblance to the shown signal and indicating the sequence of phonemes, syllables or words that corresponds to this reference signal. The recognition problem is solved by using the dynamic programming method. The learning consists in evaluating the parameters of the grammars: alphabet of elementary reference signals, acoustic-phonetic transcriptions of words and so on. Some algorithms for speech recognition are proposed: recognition of words, recognition of connected speech composed from the words of a finite vocabulary, speech recognition by phonemes without vocabulary restriction, recognition of words and phrases by phonemes. Good experimental results were obtained for all algorithmes.

Generative grammars and dynamic programming in speech recognition with learning

Initial word reference signals are built from the finite set of the reference elements and written by means of acoustic-phonetic transcription. The transformations of the initial word references permitting to obtain reference signals of words with the different rate of pronunciation are introduced. Reference phrase signals are composed of reference word signals with the help of special rules. The recognition of a given signal consists in location of the reference signal of a maximum likelihood for the given one and in indication of a word or a word sequence which composes it. Learning algorithm, which finds the transcriptions of words and the set of reference elements allows to adapt to some announcer and to change some part of the dictionary or the whole one. Described method of a recognition gives 1% of errors, 3% of refusals for 300 words. The word recognition in connected speech gives 3% of errors and 7% of refusals.

T. Vintsiuk

Papers

Generative grammars and dynamic programming in speech recognition with learning

Phoneme-by-phoneme recognition of speech composed of the words of given vocabulary