Speaker Detection on Telephone Calls Using Fusion between SVMs and Statistical Measures

This paper focuses on automatic speaker detection and identification, which is considered as the procedure of detecting and identifying the active speaker in multi-speaker conversations. That is, an automatic detection system is proposed for the task of speaker mining in telephonic conversations. This new detection system is based on an interlaced segmentation algorithm which we called ISI (Interlaced Speech Indexing) and employs two types of classifiers: Support Vector Machines and statistical measures of similarity. The experimental evaluations are conducted on a real telephonic database composed of 28 recordings, each recording contains 1, 2, 3, 4 or 5 speakers speaking sequentially, the duration of each file is between 40 s and 50 s. The proposed system uses the MFSC (Mel Frequency Spectral Coefficients) features, which are extracted from the different speech segments. Furthermore, a fusion architecture is proposed and employed to enhance the results obtained by each classifier alone. Results show that the proposed approach is interesting in speaker detection.