Talker Identification Using Reverberation Sensing System

Abstract

In this paper, we propose a robust talker identification (TI) system that can identify speakers in different reverberant environments. We describe a reverberation sensing system (RSS) that determines the approximate reverberation level of the surrounding environment, and then selects a TI engine trained in a similar reverberant environment; greater TI accuracy is achieved when training and test environments are similar. In this system, there are six TI engines trained in five reverberant environments and one nonreverberant environment. Performance is evaluated using new twelve test environments, which are all within a reverberation time (RT60) range of 0 to 1.8 seconds.The TI system using the RSS achieves average identification accuracy of 95.5%; an enhancement of 9% over the TI engine trained in a non-reverberant environment.

Modulation Spectral Features for Robust Far-Field Speaker Identification

“…ASI is based on the average log-likelihood measure computed for active speech frames (9) where denotes the 25-dimensional MFCC feature vector and the GMM parameters obtained for speaker . Given a group of speakers, the identified speaker is obtained using the following log-likelihood test (10)…”

“…During testing, a room impulse response classifier is used to determine which speaker model to use. Similarly, in [9] six models are used per speaker to represent unreverberant and five levels of reverberant speech (ranging from low to high). For testing, a "reverberation sensing system" is used to decide which speaker model to use.…”