Performance evaluation of software for the spectral analysis of speech signals in a MIPS based architecture

Abstract

Speech recognition and algorithms for audio encoding/decoding are large and complex. Embedded systems tend to have limited resources, so in order to develop efficient speech analysis applications for these platforms, it is important to evaluate the performance of speech processing algorithms. This paper presents the performance evaluation of an application for speech signals analysis implemented in an embedded system based on the XBurst jz4740 processor, which has MIPS based instruction set architecture (ISA). Two versions of a speech signal analysis application were designed using two algorithms for the spectral data extraction: Fast Fourier Transform (FFT) and Linear Predictive Coding (LPC). The two versions were implemented in the embedded system. Finally, a performance evaluation of the two versions implemented on the embedded system is carried out, measuring the response time, memory footprint and throughput. The results show that the response time is less than 10 seconds for speech signals with less than 214 samples, and the memory footprint is less than 25% of the maximum capacity. For larger signals, the system reduces its performance and it reaches memory saturation for signals with around 216 samples.