| Project/Area Number |
17500133
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Perception information processing/Intelligent robotics
|
|---|
| Research Institution | National Research Institute of Police Science |
|---|
Principal Investigator |
OSANAI Takashi National Research Institute of Police Science, Department of Fourth Forensic Science, Chief, 法科学第四部, 室長 (70392264)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
OZEKI Kazuhiko University of Electro-Communications, Faculty of Electro-Communications, Professor, 電気通信学部, 教授 (50214135)
KAMADA Toshiaki National Research Institute of Police Science, Department of Fourth Forensic Science, Researcher, 法科学第四部, 研究員 (10356173)
MAKINAE Hisanori National Research Institute of Police Science, Department of Fourth Forensic Science, Researcher, 法科学第四部, 研究員 (20415441)
|
|---|
| Project Period (FY) |
2005 – 2006
|
| Project Status |
Completed (Fiscal Year 2006)
|
| Budget Amount *help |
¥3,600,000 (Direct Cost: ¥3,600,000)
Fiscal Year 2006: ¥1,600,000 (Direct Cost: ¥1,600,000)
Fiscal Year 2005: ¥2,000,000 (Direct Cost: ¥2,000,000)
|
|---|
| Keywords | speaker recognition / prosody / speaking style / fundamental frequency / intonation / feature parameter transformation / individuality / forensic science / アクセント |
|---|
| Research Abstract |
Speaker recognition have mainly used acoustic feature extracted from spectral envelope of speech sounds. The feature, which is related to vocal tract, is not intentionally affected by speakers. Therefore, this feature is widely used on speaker recognition, but this is easily to be affected by a transmission characteristic. In recently, there are used prosodic features such as pitch, because there are hard to be affected in environment with many noises. In this research, we performed three studies, which studies relation to fundamental frequency and the speaking style and a study for improvement of speaker recognition, as follows.
1. The relation to fundamental frequency with speaking style.
We recorded voices that fundamental frequency, loudness and speech rate are different, were spoken by many speakers. We compared distributions of fundamental frequency with speaking styles. As results, these frequency distributions were different with speakers and speaking styles, but a difference of
… More
these distributions which were normalized by its average fundamental frequency tended to become small.
2. Text-dependent speaker verification by using DP trace as speaking style.
For focusing on speaking styles, we examined an effect of using the DP trace information on the dynamic time warping. As results, it was shown that speaker verification was possible using just DP trace information, but speaker verification rate was lower than conventional technique to compare feature parameter.
3. Feature parameter translation for speaker recognition.
We propose a feature parameter transformation method to improve the accuracy of speaker verification. The transformation is performed in two stages. In the first stage, we standardize a parameter by subtracting the average, and then dividing it by the standard deviation. In the second stage, we normalize the parameter by the norm of each feature vector. The results of the experiments using vowels uttered in isolation showed an approximately 3-point improvement in the average speaker verification rate after applying the transformation. Less
|
