A Self-Organizing Neural Network Architecture for Auditory and Speech Perception with Applications to Acoustic and other Temporal Prediction Problems.
Abstract
This project is developing autonomous neural network models for the real-time perception and production of acoustic and speech signals. The models have disclosed a common mechanism of nonlinear resonance that attentively reorganizes and groups acoustic data while suppressing unexpected noise. The SPINET pitch model transforms acoustic input into a spatial map of pitch whose properties simulate the key pitch data. SPINET was embedded into an ARTSTREAM model for auditory scene analysis that separates multiple sound sources from each other. The model groups frequency components based on pitch and spatial location cues into different streams. The model simulates psychophysical grouping data, such as frequency grouping across noise or ear of origin. These resonant streams input to an ARTPHONE model for variable-rate speech categorization. Computer simulations quantitatively generate experimentally observed category boundary shifts for VC-CV pairs, including why the interval to hear a double (VC1-C%V) stop is 150 msec longer than that to hear two different stops (VC1-C2V). This model uses resonant feedback between list categories and an automatically gain-controlled working memory. (AN)
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1995
- Accession Number
- ADA298051
Entities
People
- Stephen Grossberg
Organizations
- Boston University