Human-like dynamic programming neural networks for dynamic time warping speech recognition

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Speech Processing

Background:

• Traditional speech recognition methods face challenges with pattern variability.
• Human auditory processing offers insights into robust speech recognition.

Purpose of the Study:

• To present a human-like dynamic programming neural network (DPNN) for enhanced speech recognition.
• To analyze the dynamics and properties of the proposed DPNN architecture.
• To demonstrate the superiority of the DPNN method over conventional approaches.

Main Methods:

• Developed a DPNN inspired by human cognitive processes for pattern correlation.
• Configured neural networks where minimum energy states represent optimal pattern matching.
• Analytically explained the dynamics and properties of the neural network.
• Performed simulations for speaker-dependent isolated word classification (0-9, A-Z).

Main Results:

• The DPNN achieved superior performance in classifying speaker-dependent isolated words compared to conventional methods.
• Simulations demonstrated the effectiveness of the network's energy function for pattern correlation.
• The method showed promise for practical application with hardware implementation.

Conclusions:

• The human-like DPNN offers a more effective approach to speech recognition.
• The analytical framework provides a basis for understanding and improving DPNNs.
• The method's efficiency and accuracy suggest potential for real-world speech technology.