An undergraduate Mandarin speech database for speaker recognition research

This paper describes the development of a new speech database for speaker recognition research, UMSD (undergraduate Mandarin speech database). In UMSD, there are total 12 sessions of utterances for each of the selected 24 undergraduate students, while all recordings are conducted in different session intervals. The phonetically balanced corpus content include isolated digits (0~9), digit strings (5 phone numbers and 2 postal codes), words and phrases with different length from 1 to 10 characters (10 for each given length), the Chinese Phonetic Alphabet Table (21 Initials and 35 Finals), 2 ancient poems and a 200 words paragraph extracted from a well-known essay. Additionally, in order to effectively extract and process the interesting speech segments from UMSD, a speech database management system has been proposed on the base of MATLAB and MS-ACCESS. Results of preliminary evaluation show that the performance attained with UMSD is good, it not only meets the needs of our own recent effort in text-dependent and text-independent speaker recognition, but also allows the further research of the long term intra-speaker variability thanks to its multi-session records with different session intervals.