Definitely Not Pattern Matching:- A Method in Automatic Speech Recognition

A technique for use in automatic speech recognition (ASH) is reported which does not employ traditional pattern matching techniques.The ti-ieoretioai basis derives Irom Popper's theory oi the growth oi scienti c knowledge, and claims to solve the problem at applying highly abstract theoretical knowledge to the problem oi speech recognition.This problem lies with the construction cl bridging relations between theoretical and observation languages.Existing ASR systems attempt to express abstract knowledge in observation language terms.Observation languages are incapable at describing abstract theories.The new technique Is a combination oi recognition heuristic and researdt methodology.A detailed description oi the technique Is presented lit the term ot a commentary on Its application to the detection at the voicing pulse.Stenlng with a trivial universal theory ol the voicing pulse.constraining conditions are derived which exclude all hypotheses which tell to conionn to the theory.The tests are implemented as a Prolog program and the results shown.No comparisons are made by the program between prototypical patterns.A new theory Is titan constructed to take account oi the de ciencies In the old one and the process repeated.It Is demonstrated that with the growth oi knowledge an Increasineg greater recognition accuracy Is achieved. PROBLEMWe may consider the problem oi speech recognition is as Iollows.A speaker s vocal tract produm various energies.What we require.Is that by perionnlng various measurements upon the acoustic (and possibly other) energies.a computer system should determine whether vocal tract activity is taking place and it so.determine the nature of that activity.Given that a suf cient accuracy Is obtained.the system could be used as a Speech Input lnteriace to a computer application programme.We distinguish between the problem of recognising what aspeaker 'seys' In the sense oi what he articulates.and the problem oi Interpreting what he says.Note that.from this perspective.recognising what a speaker says is a problem oi physics and not at psymology.Thus we are interested In programming a machine to deterrnlne the Identity oi speci c physical phenomena as they mr In Its environment.A second problem is deten-nlning suitable two way communication protocols Involving enor protection and usability Issues.In this paper we shall only be dealing with the problem at determining what speech activity is taking place In the recognisers environment.speci cally. of detennlning the presence at vobing pulses.We are not concerned with recognising a pattern.We are concerned with recognising speci c physical phenomena as they occur In the ' environment.which Is a diiierent problem.'Our approach starts from the premise that every human vocal trad has the same bask: design but ls sghtly dliierent from thatoi any other.possibly lor the reason that It facilitates speaker identi cation.We take It as fundamental that these ditlerences will be due to relatively minor variations on that basic design.There should.then.exist a number at universal generallsatlons which may be made about the way In which human vocal tracts interact with their environment such that ior any speaker.it Is possible to deduce what heis saying.In the sense identified.from the