Data-driven Based Speaker Selection for Acoustic Modeling

【摘要】：正This work item intends to investigate speaker selection techniques for ASR AMs training and adaptive training.For some special users,their performances of ASR are bad.To solve the problem,usually some speech data of the users need to be collected for adapting the AMs.However it's always difficult to get enough data of the target speakers in real applications.To increase the adaptation data,we can use some data from other speakers.To ensure the borrowed data are useful to improve the performance of AMs,we'd better select speech data that are similar to those of target user.Speaker selection technique will be used to solve the problem.Our experiments showed that about 25%WER(Word Error Rate) can be reduced by the proposed method.