Language identification using spectro-temporal patch features

We present a novel approach for automatic Language Identification (LID) using spectro-temporal patch features. Our approach is based on the premise that speech and spoken phenomena are characterized by typical visible patterns in timefrequency representations of the signal, and that the manner of occurrence of these patterns is language specific. To model this, we derive a randomly selected library of spectro-temporal patterns from spoken examples from a language, and derive features from the correlations of this library to spectrograms derive from the speech signal. Under our hypothesis, the relative frequency of correlation peaks must be different for different languages. We model this by learning a discriminative classifier based on these features to detect the presence of the language in a recording. The proposed approach has been tested on two different datasets: the VoxForge multilingual speech data and the LDC2005S26 corpus. Preliminary results indicate that our proposed approach can achieve an accuracy of 85-93%, and perform significantly better than a non-phonetic HMM-based classifier.