Identification of aged speakers with speaker recognition techniques and its improvement

摘要

The population of aged people is getting increased these days. In this situation, spoken dialogue systems user-friendly to aged speakers as well as young ones are expected to be required. A function needed to realize the high user-friendliness of the systems is to estimate speakers' age by hearing their voices, and change the system's behaviors adequately. This paper firstly describes a listening test where subjects were asked to estimate age of speakers subjectively only by hearing their voices and judge whether they needed special care about speech communication. After the test, we carried out a series of experiments where we tried to distinguish automatically between the aged speakers and the others. In the first experiment, speaker recognition techniques were immediately used and 91 investigated other acoustic features to improve the performance. Experiments showed the validity and usefulness of speech rate and local power perturbation for the improvement. After all, the performance was raised to be 95%.