Research On Time-varying Robustness In Speaker Recognition

Voiceprint recognition also is called Speaker recognition. It is one of biology recognition. Since it is put forward, there are so many researcher think whether recognition rate has changed with time. The research found that the recognition rate has changed with time. This paper research on time-varying robustness in speaker recognition. Its main work include:1. We built a voiceprint databases that can solve the problem which is about speaker recognition time-varying robustness. From the database that we know, we can find when the time goes by, the rate of speaker recognition will become lower. Therefore, we combine the practical application, and design the database as this: firstly, interval time is a week; secondly, interval time is equal; thirdly, text content is text independent.2. We have analysis the speaker characteristic with time-varying. In order to research the speaker characteristic with time-varying, we need to reduce the influence from the other cause. So, we choose some voice which come from the same man and the different week. Then we process the voice characteristic by DTW, and find it changing with linear. At last we based on law of linear propose some method to solve the problem of time-varying robustness in speaker recognition.3. We propose a method that is through updating the model to improve the rate of speaker recognition. Based on time axis displacement direction, and constantly update data of training model, remove the corresponding time remote data at the same time, and constantly update training model, adopt the method of updating model to improve robustness is the most direct and most effective way, this method can keep advancing with The Times, has a good effect, but also a large amount of data storage problem.4. Voiceprint recognition of time-varying problem, to some extent can be seen as the change of the channel, according to the principle of the happening of the people, people change is mainly due to the track changes, and the change of the channel can be thought of as the change of the channel, in solving the problem of channel change, on the speech recognition has been using Joint Factor Analysis(be Factor Analysis, JFA), as well as the further development of the I- vector technology. So this article use the i-vector technology uses of time-varying problems in speaker recognition.5. In the use of i-vector model of the premise, according to the characteristics of time-varying characteristics, is put forward based on probability and Linear Discriminant Analysis(Probabilistic Linear Discriminant Analysis, PLDA) solution. Due to the characteristics of changes linearly with time, so that you can learn by PLDA time-varying characteristic of the linear change, so as to solve the problem of time-varying robustness.