Query-by-example spoken term detection using bessel features
Drisya Vasudev, Suryakanth V. Gangashetty, K. K Anish Babu, K S Riyas · 2015
In this paper, a new set of features for addressing the problem of unsupervised query-by-example spoken term detection is proposed. The main purpose of this is to find a spoken query in large speech databases. In unsupervised audio search, language specific resources are not required. Thus this system is more appropriate in cases where enough training data is not available for creating an Automatic Speech Recognition(ASR). Current state-of-the-art techniques use Mel Frequency Cepstral Co-efficients(MFCC), Linear Predictive Cepstral Coefficients(LPCC) etc. as the features. For improving the performance of the system, Fourier Bessel Cepstral Coefficients(FBCC) is used in this paper. Here, from the spoken example of a keyword, segmental Dynamic Time Warping is used to compare the Gaussian Posteriorgrams, which are created from the FBCC feature vector. The keyword detection result obtained using MediaEval 2012 database shows that this system outperforms the one that uses MFCC alone due to the fact that Bessel features are more efficient for representing speech signals compared to MFCC.