Audio Feature Extraction on SIBI Dataset for Speech Recognition

Mel Frequency Cepstral Coefficients has been regarded as the standard method of feature extraction for Automatic Speech Recognition (ASR) systems for the last few years. Its performance may be affected by multiple variables, such as the number of features, audio channels, filter width, or the types of filter banks used. In this paper, several comparisons were made to find the best combination of variables that provides the best results on the SIBI (Indonesian Sign Language) dataset, which consists of utterances of sentences by both Deaf and Hard of Hearing (DHH) and non-DHH people. Based on this experiment, although generally the ASR on DHH dataset is lower than those of the non-DHH dataset, the results are still relatively high, around 4.71 % WER and 10.30% SER compared to 0.15% and 0.40% in WER and SER, respectively.