Efficient speech edge detection for mobile health applications
Dingkun Du, Kofi M. Odame · 2011
Intelligent audio sensors that are continuously recording and analyzing sounds are a critical component of many emerging and future embedded applications. In these applications, the power budget is very tight, of which the analog front end consumes a major proportion. An efficient analog front end should adapt its power consumption to the instantaneous bandwidth of the audio signal of interest, instead of constantly consuming a fixed amount of power that assumes a fixed signal bandwidth. In this paper, we introduce a novel algorithm for identifying the edges of speech in the time-frequency domain, which is used to detect the instantaneous bandwidth of speech. A circuit implementation of our algorithm consumes 42.4μW of power and can extract the instantaneous bandwidth of a signal within an accuracy of 1% even in SNR conditions as low as 10 dB.