Centroid neural network with a divergence measure for GPDF data clustering.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Speech Recognition

Background:

• Continuous Density Hidden Markov Models (CDHMMs) are crucial for speech recognition.
• Efficient clustering of Gaussian Probability Density Function (GPDF) data is essential for Vector Quantization (VQ) codebook design in CDHMMs.
• Conventional methods often use only mean values for clustering, limiting accuracy.

Purpose of the Study:

• To propose an unsupervised competitive neural network for efficient GPDF data clustering.
• To introduce the Divergence-Based Centroid Neural Network (DCNN) that utilizes both mean and covariance values of observation densities.
• To enhance the performance of CDHMMs in speech recognition through improved codebook design.

Main Methods:

• Developed the Divergence-Based Centroid Neural Network (DCNN), an unsupervised competitive neural network.
• Employed a divergence measure as the distance metric within the DCNN.
• Utilized both mean and covariance values of GPDF data for clustering, unlike conventional methods that use only means.

Main Results:

• The DCNN effectively allocates code vectors, concentrating them in dense GPDF data regions and sparsifying them in sparse regions.
• When applied to Korean monophone recognition, the DCNN reduced the number of GPDFs by 65.3% while maintaining recognition accuracy.
• Comparative experiments with divergence-based k-means and self-organizing map algorithms demonstrated the DCNN's superior performance.

Conclusions:

• The proposed DCNN offers an efficient and effective method for clustering GPDF data in CDHMMs.
• The DCNN's ability to leverage both mean and covariance information leads to significant codebook size reduction without compromising accuracy.
• This approach holds promise for advancing speech recognition technologies by optimizing underlying statistical models.