A Landmark-based Model of Speech Perception: History and Recent Developments

This paper traces some of the history of the development of a model for speech perception in which words are assumed to be represented as sequences of bundles of binary distinctive features. In the model, probability estimates for feature values are derived from measurements of acoustic attributes in the vicinity of acoustic “landmarks.” Landmarks are detected based on amplitude changes in various energy bands, and landmarks or pairs of landmarks provide evidence for the existence of feature bundles. This paper discusses data and thought processes that prompted four significant changes in the formulation of the model over the past decade: rule-generated changes in segments versus modifications of cues for features, landmark detection, principles of cue selection, and the role of analysis-by-synthesis in verifying word hypotheses. The model currently allows for refinement of a cohort of words alongside the landmark-to-feature estimation process. In this view, initial estimates of some features based on local analysis of the signal are used to propose a cohort. The words in the cohort are used to inform landmark/feature estimation based on a more extended context. This process iterates between improving feature estimation and refining the cohort to arrive at the model output. INTRODUCTION This paper provides a brief review of the history of the development of a model of lexical access and discusses data and thought processes that prompted changes in the original formulation of the model. The model incorporates methods of signal processing which are derived from knowledge of relations between articulation, acoustics, and perception (the Lexical Access Project). The Lexical Access Project at MIT began in the early 1990’s with the assumption that the lexicon is represented in terms of sequences of “segments,” or bundles of binary distinctive features (Jakobson, et al., 1952), where a change in the value of a distinctive feature potentially leads to a new word. These features are grouped into two classes: articulator-free features and articulator-bound features (Table 1). Articulator-free features are not associated with a particular articulator (Halle,1992) but are associated with general characteristics of constrictions within the vocal tract and the corresponding acoustic patterns formed by these constrictions. These features classify segments into broad classes which might be described as vowels and some general classes of consonants. Articulator-bound features specify which articulators are active during production, and how these articulators are shaped or positioned. Slifka et al. Lexical Access Project From Sound to Sense: June 11 – June 13, 2004 at MIT C-86 One of the fundamental assumptions of the model is that there are locations or landmarks in the acoustic signal which are particularly rich in information related to the configurations of the larynx and of the various components of the articulatory tract. This information can be expressed in terms of acoustic cues to the distinctive features. Landmarks occur at times of a) consonant closures, b) consonant releases, c) glide minima and d) vowel maxima. Landmarks are thought to provide initial information about the manner of articulation and the order and number of intended segments, and create a road-map for subsequent acoustic analysis (Stevens, 2002). Because these locations are likely to be acoustic discontinuities or distinct minima/maxima in energy, they should be robustly identifiable to both a listener and an algorithm. More importantly, they bear a privileged relationship to the underlying segments of the utterance, in the sense that each landmark locates a region in the signal where the cues for the various features tend to be concentrated. Other acoustic discontinuities, such as the onset of voicing or closure/opening of the velopharyngeal port, also provide useful information, but do not have the same privileged relationship to underlying contrastive segments (Stevens,2002). Much of the research on this project in the 1990’s and early 2000’s was concerned with the development of methods for discerning (1) landmarks and (2) cues for the identification of articulator-bound features by examining the signal in the vicinity of these landmarks. The goal was to automate this process, and to derive from these cues the sequences of bundles of distinctive features that can then be used to access words and word sequences from the lexicon. Modules for several different types of landmarks and several features were developed, including consonant landmarks, vowel landmarks, landmarks for glides, place features for stop consonants, the feature [nasal], the voicing feature, and place features for vowels. Some of these modules involved automatic procedures, and some required hand extraction of cues. The accuracy with which these various landmarks and features could be identified in these modules ranged from about 60 percent to 95 percent for running speech. Table 1. List of distinctive features for English as grouped by articulator-free and articulator-bound classes. Articulator-bound features Articulator-free features Vowel and glide Consonant Vowel Consonant Continuant Sonorant Strident High Low Back Adv. Tongue root Spread glottis Lips Tongue blade Tongue body Round Anterior Lateral Rhotic Nasal