Author(s): Anly Paul
Published in: International Journal of Engineering Research & Technology
License: This work is licensed under a Creative Commons Attribution 4.0 International License.
Volume/Issue: Volume. 6 - Issue. 08 , August - 2017
This paper presents a feature extraction algorithm for speech at different modes such ad whispered, soft, neutral, loud and shouted. Normal feature extraction methods such as LPCC, MFCC or PNCC extract cepstral components of speech and speech recognition systems build on such features provide better recognition accuracy for speech of only normal mode. But in real life, a speech recognition system that could recognize speech of only normal mode is of no use. A better method to recognize speech at whispered, normal or shouted mode is explained in this paper. Energy of the speech signal is computed for each frame and based on the energy level speech mode is categorized. Power spectrum estimated using linear prediction is used instead of power spectrum computed directly from speech signal. MFCC coefficients are calculated from the estimated power spectrum which gives the resulting features for speech recognition at different modes. GMM, GMM-UBM or DNN can be used in training and testing phase of the speech recognition system. This will result in a proper speech recognition system working in any speech mode. The effectiveness of the approach can be demonstrated using a speech recognition system for women safety which is trained and tested with shouted speech.
Number of Citations for this article: Data not Available
7 Paper(s) Found related to your topic:
Publish your Ph.D/Master's Thesis Online