Robust speech recognition based on joint model and feature space optimization of hidden Markov models

Area of Science:

• Speech Recognition
• Machine Learning
• Signal Processing

Background:

• Robust speech recognition is challenged by various mismatch conditions between training and testing data.
• Hidden Markov Models (HMMs) are widely used for speech recognition but are sensitive to these mismatches.
• Artificial Neural Networks (ANNs) offer inversion capabilities, providing a basis for HMM adaptation.

Purpose of the Study:

• To propose and evaluate a novel HMM inversion algorithm for robust speech recognition.
• To investigate the integration of HMM inversion with model space optimization techniques.
• To demonstrate the effectiveness of joint model and feature space mismatch compensation.

Main Methods:

• Developed gradient-based and Baum-Welch HMM inversion algorithms, viewing HMMs as ANNs.
• Integrated HMM inversion with model space optimization techniques like MINIMAX.
• Applied joint model and feature space mismatch compensation strategies.

Main Results:

• The proposed HMM inversion algorithms successfully compensate for mismatch conditions.
• Joint space mismatch compensation outperforms single-space compensation methods.
• An approximate 10-dB SNR gain was achieved in low SNR environments.

Conclusions:

• The HMM inversion algorithm offers a robust approach to speech recognition under mismatch.
• Joint model and feature space compensation is a highly effective strategy for improving performance.
• This technique significantly enhances speech recognition in noisy conditions.