|Rashmi Kethireddy, Sudarsana Reddy Kadiri, Santosh Kesiraju, Suryakanth V. Gangashetty|
In this paper, we propose to use novel acoustic features, namely zero-time windowing cepstral coefficients (ZTWCC) for dialect classification. ZTWCC features are derived from high-resolution spectrum obtained with zero-time windowing (ZTW) method, and were shown to be useful for discriminating speech sound characteristics effectively as compared to a DFT spectrum. Our proposed system is based on i-vectors trained on static and shifted delta coefficients of ZTWCC. The i-vectors are further whitened before classification. The proposed system is compared with i-vector baseline system trained on Mel frequency cepstral coefficient (MFCC) features. Classification results on STYRIALECT database (German) and UT-Podcast (English) database revealed that the system with proposed features outperformed aforementioned baseline system. Our detailed experimental analysis on dialect classification shows that the i-vector system can indeed exploit high spectral resolution of ZTWCC and hence performed better than MFCC features based system.