Vector quantization of speech frames based on self-organizing maps

Area of Science:

• Signal Processing
• Artificial Intelligence
• Machine Learning

Background:

• Speech compression is crucial for efficient data transmission and storage.
• Vector quantization (VQ) is a widely used technique for data compression.
• Traditional VQ methods often require complex codebook generation.

Purpose of the Study:

• To propose a novel speech compression technique leveraging unsupervised learning.
• To implement a vector quantizer using a neural network.
• To explore the use of a self-organizing Kohonen map for codebook generation.

Main Methods:

• Speech signal processing fundamentals are reviewed.
• Vector quantization principles are discussed.
• A 2-dimensional self-organizing Kohonen map is employed to create the VQ codebook.
• Neural network-based unsupervised learning is utilized for codebook generation.

Main Results:

• Simulation results provide insights into network topology, initialization, and training strategies.
• The impact of codebook size on compression performance is analyzed.
• Speech quality is evaluated based on the generated codebook.

Conclusions:

• The proposed neural network-based vector quantization offers an effective approach to speech compression.
• The self-organizing Kohonen map provides a viable method for generating efficient speech compression codebooks.
• The technique demonstrates competitive speech quality when compared to established compression algorithms.