Application of GMMs to Speech Recognition using very short time series
Souad Friha, N. Mansouri, Hichem Arioui, Rochdi Merzouki, Hadj Ahmed Abbassi · AIP conference proceedings · 2008
This paper reports on some recent results in speech recognition using the state of the art GMMs modeling. This is done over a reconstructed multidimensional attractor that is obtained via an embedding procedure into a phase space.The novelty is being the use of very short time series of 20 ms of speech for both the training data base and the test samples.Classification accuracies reached 75.99% when four phoneme classes are concerned and 100% when there are only two phoneme classes. Experiments over two monosyllabic words gave an accuracy of 89.55%. Application of GMMs to speaker recognition, without using the traditional MFCC parameters was performed too and has resulted in an accuracy of 68.51%.