Modelling representations in speech normalization of prosodic cues.

Area of Science:

• Psycholinguistics
• Speech Perception
• Computational Auditory Processing

Background:

• The lack of invariance problem highlights challenges in understanding speech due to speaker variability.
• Previous research has not fully explored the role of mentally stored distributional information in speech normalization.

Purpose of the Study:

• To investigate the contribution of distributional information to the normalization of prosodic cues in speech perception.
• To propose and validate a novel probabilistic parametric representation (PPR) model for speech normalization.

Main Methods:

• Modeling acoustic cue distributions from a speech corpus.
• Conducting three experiments with naturally produced and manipulated lexical tones.
• Employing advanced statistical techniques to analyze the effects of distribution parameters on normalization and identification.

Main Results:

• Distribution parameters significantly influence speech normalization and identification.
• The proposed probabilistic parametric representation (PPR) integrates speaker distribution and indexical information effectively.
• PPR remains accessible during speech perception, even with contextual information present.

Conclusions:

• Mentally stored distributional information plays a crucial role in normalizing prosodic cues.
• The probabilistic parametric representation (PPR) offers a new framework for understanding speech normalization.
• This model accounts for how listeners process speech from both familiar and unfamiliar talkers, with or without context.