Related Experiment Videos
A modular and hybrid connectionist system for speaker identification
1C.N.R.S., L.I.P.N. URA-1507, University of Paris-Nord, Villetaneuse, France.
Neural Computation
|July 1, 1995
Summary
This study introduces a modular/hybrid system for speaker identification, combining connectionist and Hidden Markov Model modules. The novel approach achieved perfect speaker identification in tests, demonstrating its effectiveness for complex tasks with limited data.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Speech Processing
Background:
- Speaker identification is complex, especially with limited training data.
- Modularity in connectionist systems aids in managing complexity and incorporating prior knowledge.
- Text-independent speaker identification presents unique challenges.
Purpose of the Study:
- To present and evaluate a modular/hybrid connectionist system for speaker identification.
- To test the efficacy of modularity in complex, data-limited scenarios.
- To develop an architecture integrating multiple connectionist modules with a Hidden Markov Model.
Main Methods:
- Developed a hybrid architecture combining several connectionist modules.
- Integrated a Hidden Markov Model module into the system.
- Evaluated the system on the DARPA-TIMIT database with 102 speakers.
Main Results:
- Achieved perfect speaker identification in tests.
- Demonstrated the effectiveness of the modular/hybrid approach.
- Validated the system's performance on a substantial speaker population.
Conclusions:
- The modular/hybrid connectionist system is highly effective for speaker identification.
- Modularity is a valuable technique for complex AI tasks like speaker recognition.
- The proposed architecture offers a robust solution for text-independent speaker identification.