Related Experiment Video
Updated: Jul 10, 2026
05:48
Memorization-Based Training and Testing Paradigm for Robust Vocal Identity Recognition in Expressive Speech Using Event-Related Potentials Analysis
Published on: August 9, 2024
Approximated mutual information training for speech recognition using myoelectric signals.
1Department of Systens & Computer Engineering, Carleton University, Ottawa, Ontario, Canada.
Summary
A novel approximated maximum mutual information (AMMI) algorithm enhances myoelectric speech recognition accuracy. This method improves hidden Markov model (HMM) training for better speech classification compared to traditional methods.
Area of Science:
- Biomedical Engineering
- Signal Processing
- Speech Technology
Background:
- Automatic speech recognition (ASR) can be achieved using myoelectric signals from facial articulatory muscles.
- Hidden Markov Models (HMMs) are commonly used for classifying these myoelectric signals.
- The standard Maximum Likelihood (ML) training algorithm for HMMs optimizes for observation likelihood, not classification accuracy.
Purpose of the Study:
- To introduce a new training algorithm, approximated maximum mutual information (AMMI), for HMMs in myoelectric speech recognition.
- To improve the classification accuracy of facial myoelectric signals for speech recognition.
Main Methods:
- Developed and applied the AMMI training algorithm to HMMs for myoelectric speech recognition.
- Compared the performance of AMMI-trained HMMs against ML-trained HMMs.
Main Results:
- AMMI training consistently reduced error rates in myoelectric speech recognition.
- The AMMI algorithm improved classification accuracy by approximately 3% on average compared to ML training.
Conclusions:
- The AMMI training algorithm offers a significant improvement in accuracy for myoelectric speech recognition.
- Optimizing HMM parameters for discrimination via mutual information enhances speech classification performance.