Cloud Point Matching for Text-Independent Speaker Identification

摘要

In Text-Independent speaker identification, the individual that produced some captured speech signal has to be identified without his collaboration, he might not even know that he is being the subject of an identification process. The system could not ask the individual to utter some specific word or phrase, which is precisely what is done in Text-Dependent speaker recognition. Text-Independent speaker identification is far more complicated since we cannot simply measure the similarity of an utterance of a word or phrase to another utterance made by the same speaker of the same word or phrase in which case we could use the dynamics of the speech signal. In this paper we search in the speech signal looking for voiced speech segments and estimate its first three formants, so we end up with a three-dimensional point cloud for each speaker of the collection of known speakers. To identify a speaker we have to measure the similarity of a point-cloud from an unknown speaker to the point-clouds that belong to known speakers, we do that by searching for local structures in the cloud in a way that is highly scalable and robust. We performed tests with both a collection of our own in Spanish and with the English Language Speech Database for Speaker Recognition (ELSDSR) from the Technical University of Denmark achieving results that improve recent published work with ELSDSR.