A Comparison of Low-Complexity Real-Time Feature Extraction for Neuromorphic Speech Recognition

Area of Science:

• Neuromorphic Engineering
• Speech Recognition Technology
• Artificial Intelligence

Background:

• Traditional speech recognition systems often require significant computational resources and energy.
• Neuromorphic computing offers a promising alternative for energy-efficient, real-time processing.
• Spiking neural networks and silicon cochleas are key components for bio-inspired auditory processing.

Purpose of the Study:

• To develop and evaluate a real-time, low-complexity neuromorphic speech recognition system.
• To compare the performance of different feature extraction methods in terms of accuracy, energy, and memory.
• To assess the hardware implementation's performance against software simulations.

Main Methods:

• Utilized a spiking silicon cochlea for auditory feature extraction.
• Employed a population encoding method based on the Neural Engineering Framework (NEF) and Extreme Learning Machine (ELM) classifier.
• Investigated several feature extraction techniques with varying computational and memory demands.
• Validated the system using the N-TIDIGITS18 dataset.

Main Results:

• A fixed bin size feature extraction method achieved 95% accuracy in software, outperforming previous methods with reduced energy (~3x) and memory (~25x) usage.
• Hardware implementation yielded a 94% accuracy, with minor reductions attributed to hardware-specific correlations.
• Increased hidden nodes in the ELM classifier improved hardware accuracy at the expense of increased memory and energy consumption.

Conclusions:

• The proposed neuromorphic system offers a viable, energy-efficient solution for real-time speech recognition.
• Feature extraction optimization is crucial for balancing accuracy, resource usage, and hardware performance.
• Further hardware optimization can enhance accuracy, though trade-offs with resource consumption must be considered.