Summary: | 碩士 === 國立臺北科技大學 === 電腦與通訊研究所 === 97 === The most widespread approach to automatic language identification in the past has been the multiple language Phone recognizer followed by n-gram language modeling (PPRLM). This system has consistently provided good results for the task of language identification. By contrast, Gaussian mixture model (GMM) systems, which measure acoustic characteristics, are far more efficient computationally but have tended to provide inferior levels of performance.
In this thesis, we present two GMM-based approaches to language identification. The approaches include both acoustic scoring and GMM tokenization system. In the acoustic scoring system, we use shifted delta cepstra (SDC) feature to describe that additional temporal information about the speech into the feature vectors. System performance is evaluated on the NIST LRE 2009 corpus. Experimental results on 23 language recognition task showed that fusion of the proposed PPRLM、SDC GMM and GMM tokenization achieves a closed set equal error rate 17.38%.
|