Pattern Rect~.lnition Vol. 14. Nos. I 6. pp. 261 265. 1981 Pfinlcd ill Great Ilritam [)031 3203/81/070261 05 $02.00/0 Pergamon Press Ltd 1981 Pattern Rccognilion Society A MULTILEVEL APPROACH TO PATTERN PROCESSING ROGER K. M(X)RE Department of Phonetics and Linguistics, University College London, Gower Street, London WCIE 6BT, U.K. (Received 9 January 1980; receivedJbr publication 22 December 1980) Abstract - This paper presents a methodology for describing multilevel pattern processing systems. It is suggested that any pattern processor can be adequately described in terms of multiple hierarchies of two types of fundamental mechanism : (1) a process which performs the pattern recognition functions of analysis and synthesis and (2) a process which performs the syntactic functions of parsing and generation. A computer implementation of these principles is outlined which enables a range of systems to be configured. Examples of speech and non-speech pattern processing are presented. Pattern processing Descriptive technique Multilevel Analysis/synthesis Syntactic rules Speech recognition I. INTRODUCTION Since the year 1970 the field of automatic speech recognition has been concerned with the structure and organisation of complex pattern processing systems. This concern arose out of the realisation that infor- mation from many different sources had to be coordi- nated in order to recognise continuous speech. It was felt that perhaps correct understanding of a spoken phrase was more important than correct recognition of all its component words, and that this could be achieved by a mixture of bottom-up and top-down processing using suitable sources of knowledge such as acoustics, syntax and semantics: 1~ Automatic speech recognition programs which reflected these views were termed 'speech understanding systems' (SUS's) and from 1971 to 1976 several large-scale SUS projects flourished (see Klatt 12) for a review). During the same period, the author was involved with developing techniques for measuring and com- paring the performances of automatic speech re- cognisers. Initially only isolated-word recognisers were considered, [3) but later interest became focused on the more complex structures of speech understand- ing systems. The result of this latter work was not an evaluatory technique, but a methodology for describ- ing the structures of different speech understanding systems in terms of basic pattern processing mechanisms. 14~ 2. DESCRIPTIVE METItODOLOGY During the investigation of different SUS organi- sations it became apparent that there was universal concern for the key concepts of'level of representation' and 'unit of representation'; concepts which are of equal importance to the description of the human perceptual apparatus as well as to other pattern processing structures. Further study of these concepts led to the conclusion that any pattern processing system could be adequately described in terms of multiple hierarchies of two fundamental types of process: interlevel processes which relate represen- tations at one level to representations at another, and intralevel processes which relate representations at one level to other representations at the same level. The first of these processes, the interlevel process, is deemed to be the site of the pattern recognition function of matching. Strings of units at a particular level are matched against a list of equivalences to determine the appropriate unit at the next level. For example, the phoneme string [(s)0)(k)(s)] would be recognised as the lexical construct (6) using the equivalence [(s)0)(k)(s)] = (6). Such a conversion from a low-level sub-construct string to a high-level construct is termed 'analysis'. The alternative con- version from construct to string is termed 'synthesis'. The list of equivalences constitutes a dictionary, hence an interlevel process is referred to as the dictionary 'look-up' or LUP process. In pattern recognition terms the dictionary performs as a training set. Fig. l(a) shows the graphical notation for a LUP process. The second fundamental type of process, the in- tralevel process, is the site of sequential syntactic behaviour. Units are collected to form strings which are parsed such that the units in the string conform to some specified relationship. The opposite process of breaking up a string is termed 'generation'. The sequential relationships is an intralevel process are directed by a set of syntactic rules hence it is referred to as the 'rules' or RLS process [Fig. l(b)]. Using this methodology a multilevel processing system can be viewed as an interconnected network of 261