Speech recognition informed by distinctive feature theory: The featurally underspecified Lexicon model and its implications

Abstract

We present a speech recognition engine that implements the Featurally Underspecified Lexicon calculus (FUL). The FUL model defines an inventory of privative phonological features that is necessary and sufficient to describe contrasts between phonemes in any language in the world. The model also defines conditions for comparing feature bundles recovered from the signal with segments defined in the lexicon: a feature may MATCH (the feature is present in both the signal and the lexicon), it may MISMATCH (the feature in the signal is impossible in tokens of the segment in the lexicon, e.g. when a stop-burst, which indicates the [PLOSIVE] feature, is compared with a segment carrying the [CONTINUANT] feature in the lexicon), or it may provoke a NOMISMATCH response when the feature in the signal is not part of the segment in the lexicon. This matching mechanism also accounts for asymmetries as they are observed in natural speech, as [CORONAL] assimilates to [LABIAL] but not the other way around. The engine computes distances to neighboring words according to a coherence measure to simulate co-activation in the lexicon. We will demonstrate online the operation of this engine in English and German.