An investigation of sub-band FM feature extraction in speaker recognition

Following recent evidence that FM features extracted from a sub-band decomposition of speech are highly uncorrelated, this paper investigates the effect of the number of auditory scale sub-bands in FM based front-end processing. For this study, a newly developed robust FM extraction method based on the least square differential ratio is used to extract features, comprising one FM component per sub-band. Automatic speaker recognition experiments were conducted on the cellular NIST 2001 database, with the number of filters in the front-end varied from 6 to 26. Performance degradation was observed for very low numbers of filters and very high numbers of filters. Results show that for a 4 kHz speech bandwidth, a minimum of 10 and a maximum of 18 sub-bands is a suitable choice for speech front-end applications such as automatic speaker recognition.