Efficient FPGA implementation of a knowledge-based automatic speech classifier

Speech recognition has become common in many application domains, from dictation systems for professional practices to vocal user interfaces for people with disabilities or hands-free system control. However, so far the performance of Automatic Speech Recognition (ASR) systems are comparable to Human Speech Recognition (HSR) only under very strict working conditions, and in general far lower. Incorporating acoustic-phonetic knowledge into ASR design has been proven a viable approach to rise ASR accuracy. Manner of articulation attributes such as vowel, stop, fricative, approximant, nasal, and silence are examples of such knowledge. Neural networks have already been used successfully as detectors for manner of articulation attributes starting from representations of speech signal frames. In this paper an optimized digital Knowledge-based Automatic Speech Classifier for real-time applications is implemented on FPGA using six attribute scoring Multi-Layer Perceptrons (MLP). Digital MLP key features are a virtual neuron architecture and use of sinusoidal activation functions for the hidden layer. Implementation results on FPGA show that use of sinusoidal activation functions decrease hardware resource usage of more than 50% for slices, FFs, LUTs and more than 35% for FPGA RAM blocks when compared with the standard sigmoid-based neuron implementation. Furthermore, neuron virtualization allows for a significant decrease of concurrent memory access, resulting in improved performance for the entire attribute scoring module.