Neural Networks-based Automatic Speech Recognition for Agricultural Commodity in Gujarati Language

In this paper, we present a development of Automatic Speech Recognition (ASR) system as a part of a speech-based access for an agricultural commodity in the Gujarati (a low resource) language. We proposed to use neural networks for language modeling, acoustic modeling, and feature learning from the raw speech signals. The speech database of agricultural commodities was collected from the farmers belonging to various villages of Gujarat state (India). The database has various dialectal variations and real noisy acoustic environments. Acoustic modeling is performed using Time Delay Neural Networks (TDNN). The auditory feature representation is learned using Convolutional Restricted Boltzmann Machine (ConvRBM) and Teager Energy Operator (TEO). The language model (LM) rescoring is performed using Recurrent Neural Networks (RNN). RNNLM rescoring provides an absolute reduction of 0.69-1.18 in % WER for all the feature sets compared to the bi-gram LM. The system combination of ConvRBM and Mel filterbank further improved the performance of ASR compared to the baseline TDNN with Mel filterbank features (5.4 % relative reduction in WER). The statistical significance of proposed approach is justified using a bootstrap-based % Probability of Improvement (POI) measure.