Global optimization of a neural network-hidden Markov model hybrid

Area of Science:

• Artificial Intelligence
• Speech Processing
• Machine Learning

Background:

• Artificial Neural Networks (ANNs) excel at function approximation for acoustic parameters.
• Hidden Markov Models (HMMs) are effective for modeling the temporal dynamics of speech signals.
• Integrating ANNs and HMMs offers a synergistic approach to speech recognition challenges.

Purpose of the Study:

• To explore the integration of multilayered, recurrent ANNs with HMMs for speech recognition.
• To develop a unified system where ANN outputs serve as observation vectors for HMMs.
• To propose a global optimization algorithm for the integrated ANN-HMM system parameters.

Main Methods:

• Utilizing multilayered and recurrent Artificial Neural Networks (ANNs) for acoustic feature extraction.
• Employing Hidden Markov Models (HMMs) to capture the sequential nature of speech data.
• Developing a novel algorithm for the global optimization of all system parameters.
• Conducting speaker-independent recognition experiments on the TIMIT continuous speech database.

Main Results:

• The integrated ANN-HMM system demonstrated effective performance in speaker-independent speech recognition tasks.
• The approach successfully combined the strengths of ANNs in acoustic modeling and HMMs in temporal modeling.
• Experimental results on the TIMIT database validated the proposed integration and optimization strategy.

Conclusions:

• The integration of ANNs and HMMs provides a powerful framework for advanced speech recognition.
• The proposed global optimization algorithm is effective for tuning the combined system parameters.
• This hybrid approach shows significant promise for improving the accuracy and robustness of continuous speech recognition systems.