Research On Classification Of Speaker’s Properties Based On Multi-instance Learning

The classification of speakers’ properties is a process of estimating the speakers’geographical and gender information based on speech. It has important applicationvalue in many fields, for example: processing of multilingual information,machine translation, criminal investigation information of publicsecurity, military intelligence gathering and so on.Multi-instance learning is an effective machine learning algorithm to solve theambiguity problem. It is commonly used image retrieval,text classification and someother static pattern classification, but not commonly used in speech signal processing.Multi-Instance Learning, a time-varying method of gender identification is proposed inthis paper. What’s more, it’s applied to the dialect identification. The main achievementsare as follows:1.The Chinese dialect database was extended and labeled mainly on thenorthern dialect: min dialect, xiang dialect, gan dialect, wu dialect, yue dialect,kejiadialect and mandarin. Each segment of the speech was labeled with the speakers’information, such as gender, age, record of time and city.2. Time-varying multi-instance model was proposed. Due to the continuity of thespeech signal, the speech segments of the speech signal were cut into several piecesmanually, acoustic characteristics was extracted from speech signal, finally K-meansalgorithm was used to get instances from bags.3. Two-point model was proposed to replace the single point model. Underdifferent scale transformation, different categories of maximum diversity density pointwere calculated respectively.4. Bags-kNN classification was proposed. On the backend stage, distancemeasure was solved the problem between sets, the traditional threshold judgementwas replaced to enhance the performance of the classifier.