The Study Of Age Automatic Estimation Based On Gender Pre-classification

Age estimation is a technology that judge one’s age automatically using an age estimation system on the computer. The technology has broad application prospects in many fields, for example, information retrieval, man-machine communication, criminal investigation.At present, most systems used in multiple age-grouping estimation are consisted of single speech feature or single classifier. Although much work about feature extraction and system design has been done, some problems such as the instability of feature and the low classification accuracy of single-tier system are still unresolved. What’s more, there’s lacking of an evaluation platform that is the recognized age speech corpus. For those problems, this paper study up on database building, feature extraction, classification identification and has made the following innovative achievements.1. Set up an age speech databaseReferring to design standards of speech database in the world, we improved the distribution of speaker’s age,gender and eventually set up an age speech database which contains three kinds of ages and be equally male and female. Besides, the information about the speaker of each speech segment such as gender, age is labeled which is helpful for the expansion of the database function.2. Set up an age estimation system fusing the gender pre- classification systemNow an age estimation system is usually designed with single feature and single classifier. As a result, the classification accuracy of system is often on the low side. In this paper gender classification is conducted first. According to the hierarchical classification principle, the speaker is classified as a child or not. Then the young and the old are judged. Combining the characteristic of each subtask, different features such as fundamental frequency, short-term energy are used to improve the final estimation effect of system.3. Put forward a new adult gender classification method based on modifiedCitation-kNNCitation-kNN is often used in image processing. It’s the first time that imports Citation-kNN into gender classification. A new speech Multi-Instance(MI) bag generating method under GMM is introduced in this paper; then the distance measurement of Citation-kNN is modified, and the system has been simplified. The test experiments show that using modified Citation-kNN algorithm in gender classification is feasible, and the system classification effect is better than some traditional algorithms.4. Put forward an age subclass estimation system based on band weighted Melfrequency cepstrum coefficient(MFCC)Frequency band information obtained from speech signal by using discrete Fourier transform is important to age estimation task, and the contribution of each band is different. According to the F-ratio, the overall contribution degree is calculated based on frequency band energy. When computing MFCC, frequency band energy, as the output of each Mel filter, is weighted according its contribution degree. As a result, there’s different emphasis on different frequency bands. The experiment results show that modified MFCC is better than traditional MFCC in reflecting age information.