Accelerating dynamic time warping for speech recognition with SSE

Abstract

This study presents a significant enhancement to the Dynamic Time Warping (DTW) algorithm for real-time applications like speech recognition. Through integration of SIMD (Single Instruction Multiple Data) instructions to distance function, the research demonstrates how SSE accelerates DTW, markedly reducing computation time. The paper not only explores the theoretical aspects of DTW and this optimization but also provides empirical evidence of its effectiveness. Diverse dataset of 18 voice command classes was assembled, recorded in controlled settings to ensure audio quality. The audio signal of each speech sample was segmented into frames for detailed analysis of temporal dynamics. DTW search was performed on features set based on Mel Frequency Cepstral Coefficients (MFCC) and Linear Predictive Coding (LPC), combined with delta features. A comprehensive set of 27 features was extracted from each frame to capture critical speech characteristics. The core of the study involved applying traditional DTW as a baseline for performance comparison with the SSE-optimized DTW. The evaluation, focusing on computational time, included measurements like minimum, maximum, average, and total computation times for both standard and SSE-optimized implementations. Experimental results, conducted on datasets ranging from 5 to 60 WAV files per class, revealed that the SSE-optimized DTW significantly outperformed the standard implementation across all dataset sizes. Particularly noteworthy was the consistent speed of the SSE-optimized Manhattan and Euclidean distance functions, which is crucial for real-time applications. The SSE-optimized DTW maintained a low average time, demonstrating remarkable stability and efficiency, especially with larger datasets. The study illustrates the potential of SSE optimizations in speech recognition, emphasizing the SSE-optimized DTW's capability to efficiently process large datasets.