Deep Convolution Neural Network Based Speech Recognition for Chhattisgarhi

• Published in: 2018 5th International Conference on Signal Processing and Integrated Networks (SPIN) pp. 667 - 671
• Main Authors: Londhe, Narendra D., Kshirsagar, Ghanahshyam B., Tekchandani, Hitesh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.02.2018
• Subjects: Chhattisgarhi dialect; Convolution; Convolution neural network; Deep learning; Hidden Markov models; Machine learning; Mel frequency cepstral coefficient; Neural networks; Speaker independent isolated word recognition; Speech recognition; Training
• DOI: 10.1109/SPIN.2018.8474064

The existing ASR for Chhattisgarhi using conventional machine learning technique was implemented for speaker dependent speech recognition. However, the conventional machine learning based speech recognition is incapable to handle the spectral variations as well as the spectral correlation of acoustic signals. Therefore, to overcome the aforementioned limitations, authors have implemented the deep convolution neural network (DCNN) based ASR for Chhattisgarhi dialect. Unlike other deep learning models, DCNN can efficiently handle the spectral variations and spectral correlation of speech signal with the less computational burden. The experiment of isolated Chhattisgarhi word recognition was implemented on self-recorded dataset acquired from 150 subjects from various geographical parts of Chhattisgarh state. The implemented algorithm is promisingly achieving 99.49% of accuracy for isolated word recognition. The different performance paraments are presented to validate the performed experiment.