Picone: Signal Modeling ... 1Proceedings of the IEEE Second Submission: March 18, 1992Picone: Signal Modeling ... 1Proceedings of the IEEE Second Submission: May 15,1993Picone: Signal Modeling ... 1Proceedings of the IEEE Second Submission: March 18, 1992Picone: Signal Modeling ... 1Proceedings of the IEEE Final Copy: June 3,1993Signal Modeling Techniques In Speech Recognitionby,Joseph PiconeTexas InstrumentsSystems and Information Sciences LaboratoryTsukuba Research and Development CenterTsukuba, JapanABSTRACTWe have seen three important trends develop in the last five years in speech recognition. First,heterogeneous parameter sets that mix absolute spectral information with dynamic, ortime-derivative, spectral information, have become common. Second, similarity transformtechniques, often used to normalize and decorrelate parameters in some computationallyinexpensive way, have become popular. Third, the signal parameter estimation problem hasmerged with the speech recognition process so that more sophisticated statistical models of thesignal’s spectrum can be estimated in a closed-loop manner. In this paper, we review the signalprocessing components of these algorithms. These algorithms are presented as part of a unifiedview of the signal parameterization problem in which there are three major tasks: measurement,transformation, and statistical modeling.This paper is by no means a comprehensive survey of all possible techniques of signalmodeling in speech recognition. There are far too many algorithms in use today to make anexhaustive survey feasible (and cohesive). Instead, this paper is meant to serve as a tutorial onsignal processing in state-of-the-art speech recognition systems and to review those techniquesmost commonly used. In keeping with this goal, a complete mathematical description of eachalgorithm has been included in the paper.Picone: Signal Modeling ... 2Proceedings of the IEEE Second Submission: March 18, 1992Picone: Signal Modeling ... 2Proceedings of the IEEE Second Submission: May 15,1993Picone: Signal Modeling ... 2Proceedings of the IEEE Second Submission: March 18, 1992Picone: Signal Modeling ... 2Proceedings of the IEEE Final Copy: June 3,1993I. INTRODUCTIONParameterization of an analog speechsignal is the first step in the speech recognitionprocess. Several popular signal analysistechniques have emerged as de facto standardsin the literature. These algorithms are intendedto produce a “perceptually meaningful”parametric representation of the speech signal:parameters that emulate some of the behaviorobserved in the human auditory and perceptualsystems. Of course, and perhaps moreimportantly, these algorithms are also designedto maximize recognition performance.The roots of many of these techniques canbe traced to early speech recognition researchon speaker dependent technology. Today,though significant portions of speechrecognition research are now focused on thespeaker independent recognition problem,many of these parameterizations continue to beuseful. In speaker independent speechrecognition, a premium is placed ondeveloping descriptions that are somewhatinvariant to changes in the speaker. Parametersthat represent salient spectral energies of thesound, rather than details of the particularspeaker’s voice, are desired.In this paper, we will adopt a view that asyntactic pattern recognition approach tospeech recognition consists of twofundamental operations: signal modeling andnetwork searching. Signal modelingrepresents the process of converting sequencesof speech samples to observation vectorsrepresenting events in a probability space.Network searching is the task of finding themost probable sequence of these events givensome syntactic constraints. In this tutorial, wepresent an overview of popular approaches tosignal modeling in speech recognition.1.1 The Signal Model ParadigmSignal modeling can be subdivided intofour basic operations: spectral shaping,spectral analysis, parametric transformation,and statistical modeling. The completesequence of steps is summarized in Fig. 1. Thefirst three operations are straightforwardproblems in digital signal processing. The lasttask, however, is often divided between thesignal modeling system and the speechrecognition system.There are three main driving forces indesigning signal modeling systems. First,parameterizations are sought that representsalient aspects of the speech signal, preferablyparameters that are analogous to those used bythe human auditory system. This is oftenreferred to as perceptually-meaningfulparameters. Second, parameterizations aredesired that are robust to variations in channel,speaker, and transducer. We refer to this as therobustness, or invariance, problem. Finally,most recently, parameters that capture spectraldynamics, or changes of the spectrum withtime, are desired. We refer to this as thetemporal correlation problem. With theintroduction of Markov modeling techniquesthat are capable of statistically modeling thetime course of the signal, parameters thatincorporate both absolute and differentialmeasurements of the signal spectrum havebecome increasingly common.Signal modeling now requires less than10% of the total processing time required in atypical large vocabulary speech recognitionapplication. The difference in processing timebetween various signal modeling approaches isnow a small percentage of the total processingtime. The focus today has shifted towardsmaintaining high performance and minimizingthe number of degrees of freedom.Parameterizations that concisely describe theProceedings of the IEEE First Submission: July 15, 1992Picone: Signal Modeling ... 3Proceedings of the IEEE Second Submission: May 15,1993Picone: Signal Modeling ... 3Proceedings of the IEEE First Submission: July 15, 1992Picone: Signal Modeling ... 3Proceedings of the IEEE Final Copy: June 3, 1993Picone: Signal Modeling ... 3signal, can be easily computed in fixed pointhardware, and can be compressed throughstraightforward quantization techniques areoften preferred over more exotic approaches.Memory considerations often outweigh anysmall gains that may be achieved in speechrecognition performance by a new signalmodel.Historically, robustness to backgroundacoustic noise has been a major driving force inthe design of signal models. In fact, many ofthe signal models in use today were theoutgrowth of research into applicationsinvolving noisy environments: voice control ofmilitary instrumentation (speech
View Full Document
Unlocking...