Speaker Identification Based on Log Area Ratio and Gaussian Mixture Models in Narrow-Band Speech Speech Understanding / Interaction

摘要

Log area ratio coefficients (LAR) derived from linear prediction coefficients (LPC) is a well known feature extraction technique used in speech applications. This paper presents a novel way to use the LAR feature in a speaker identification system. Here, instead of using the mel frequency cepstral coefficients (MFCC), the LAR feature is used in a Gaussian mixture model (GMM) based speaker identification system. An F-ratio feature analysis was conducted on both the LAR and MFCC feature vectors which showed the lower order LAR coefficients are superior to MFCC counterpart. The text-independent, closed-set speaker identification rate, as tested on the down-sampled version of TIMIT database, was improved from 96.73%, using the MFCC feature, to 98.81%, using the LAR features.