End-to-End Topic Classification without ASR

This document explores an end-to-end model for topic classification without automatic speech recognition(ASR) system. In general, we always employ the ASR system to convert the speech recording to text and then use the standard natural language processing(NLP) knowledge to complete the topic identification task. However for low-resourced language, the lack of transcribed text and good language model results in the absence of practical speech recognition system. In this case, our paper proposes an end-to-end system for topic modeling based on mel-frequency cepstrum coefficients(MFCCs) feature. Comparing with the lexical discovery methods (such as segment dynamic time warping(DTW)), our method which can be applied to large-scale dataset which significantly reduces training time and model complexity.