Toward a model for lexical access based on acoustic landmarks and distinctive features

Area of Science:

• Speech processing and phonetics
• Computational linguistics
• Acoustic phonetics

Background:

• Traditional speech recognition models often struggle with the nuances of acoustic variability.
• Understanding the underlying distinctive features of speech is crucial for accurate language processing.
• Existing models may not fully capture the relationship between acoustic signals and phonemic contrasts.

Purpose of the Study:

• To describe a novel model for acoustic speech signal processing.
• To represent speech as a sequence of discrete segments with binary distinctive features.
• To develop a method for deriving word sequences from these feature representations.

Main Methods:

• The model processes acoustic speech signals in three steps: landmark detection, acoustic parameter derivation, and feature estimation.
• Acoustic landmarks identify articulator-free features (e.g., [vowel], [consonant]).
• Acoustic cues near landmarks estimate articulator-bound features (e.g., [lips], [high], [nasal]), considering contextual variations.

Main Results:

• The model successfully converts acoustic signals into sequences of feature bundles representing phonemic contrasts.
• It identifies both articulator-free and articulator-bound distinctive features from the speech signal.
• The model accounts for signal variability caused by enhancement gestures and overlapping articulatory movements.

Conclusions:

• This feature-based model offers a robust method for discrete speech representation.
• It provides insights into how acoustic information maps to phonemic contrasts.
• The model has potential applications in speech recognition, synthesis, and linguistic analysis.