FROM SPEECH SOUNDS TO SYMBOLIC REPRESENTATION – ABOUT NEW APPROACHES TO IMPOSE STRUCTURE ON SPEECH DATA During the last decades, phonetic and phonological knowledge about the structure of speech

During the last decades, phonetic and phonological knowledge about the structure of speech has provided a basis for the development of computational models for speech recognition, including automatic speech recognition (ASR). Since speech sounds are highly variable, the robustness of the mapping from a speech signal to its discrete symbolic representation is one of the most difficult problems in speech science. The mainstream approach to 'sound-to-symbol' mapping is based on the use of a small set of phonetic-phonologically motivated 'speech units', in combination with a statistical description of these units (obtained on a large speech corpus). In this approach, speech is considered a process that can adequately be represented by a sequence of such speech units ('beads-on-a-string' paradigm, [1]). The performance of computational models of speech recognition has shown that this 'beads-on-a-string' paradigm works reasonably well for utterances that do not deviate much from the patterns included in the training corpus ([2]). However, it is becoming clear that this conventional data-driven approach has serious limitations. Despite the use of ever-larger speech corpora, the performance of computational models of speech recognition faces a ceiling effect and falls short of human performance by an order of magnitude, particularly due to its poor capability to cope with unseen test conditions. Recently, several researchers have suggested exploring radically new approaches to address the sound-to-symbol representation. A common factor in all these new approaches is the use of sophisticated models to better impose knowledge-based structure on raw speech data. The issue of using phonological and linguistic structure is central in several lines of current research: on the role of fine phonetic details in lexical decoding ([3]), on the relation between (symbolic) context and pronunciation variation ([4]), and on the design of computational models for human speech processing ([5]). In all these research directions, the combination of statistical data-driven techniques with phonetic-phonological structure is crucial for further improvements.