Biologically-Inspired Spike-Based Automatic Speech Recognition of Isolated Digits Over a Reproducing Kernel Hilbert...

Area of Science:

• Computational Neuroscience
• Signal Processing
• Machine Learning

Background:

• Traditional speech recognition models face challenges with real-time dynamic structure quantification.
• Spiking Neural Networks (SNNs) offer potential for low-power neuromorphic computing but require advanced frameworks.
• Accurate modeling of time-series data in spoken words is crucial for robust automatic speech recognition (ASR).

Purpose of the Study:

• To present a novel real-time dynamic framework for quantifying time-series structure in spoken words using spike trains.
• To demonstrate the framework's superiority over existing methods in terms of accuracy and power efficiency.
• To explore the use of biologically-inspired spike generation and kernelized state-space models for speech processing.

Main Methods:

• Audio signals converted to multi-channel spike trains via leaky integrate-and-fire (LIF) generators.
• Spike trains mapped to a Reproducing Kernel Hilbert Space (RKHS) using point-process kernels.
• A kernelized recurrent state-space model trained with gradient descent learns spike input dynamics.

Main Results:

• The proposed framework, KAARMA, demonstrated superior performance compared to traditional Hidden Markov Models (HMMs) and state-of-the-art Spiking Neural Networks (SNNs).
• Achieved accurate and ultra-low power word spotting on the TI-46 digit corpus.
• Spike train inputs showed enhanced noise robustness in low signal-to-noise ratio (SNR) conditions compared to Mel-frequency cepstral coefficients (MFCCs).

Conclusions:

• The novel RKHS-based dynamic framework offers a powerful and efficient approach for speech processing tasks like keyword spotting.
• This biologically-inspired method provides a competitive alternative to conventional ASR techniques, especially in noisy environments.
• The framework's ability to model nonlinear dynamics and its kernel flexibility pave the way for future advancements in neuromorphic speech recognition.