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. Our SPINET pitch model was developed to take realtime acoustic input and to simulate the key pitch data. SPINET was embedded into a model for auditory scene analysis, or how the auditory system separates sound sources in environments with multiple sources. The model groups frequency components based on pitch and spatial location cues and resonantly binds them within different streams. The model simulates psychophysical grouping data, such as how an ascending, tone groups with a descending tone even if noise exists at the intersection point, and how a tone before and after a noise burst is perceived to continue through the noise. These resonant streams input to working memories, wherein phonetic percepts adapt to global speech rate. Computer simulations quantitatively generate the experimentally observed category boundary shifts for voiced stop pairs that have the same or different place of articulation, including why the interval to hear a double (geminate) stop is twice as long as that to hear two different stops. This model also uses resonant feedback, here between list categories and working memory.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 28, 1994
- Accession Number
- ADA285640
Entities
People
- Michael Cohen
- Stephen Grossberg
Organizations
- Boston University