Speech Analytics Based on Machine Learning

In this chapter, the process of speech data preparation for machine learning is discussed in detail. Examples of speech analytics methods applied to phonemes and allophones are shown. Further, an approach to automatic phoneme recognition involving optimized parametrization and a classifier belonging to machine learning algorithms is discussed. Feature vectors are built on the basis of descriptors coming from the music information retrieval (MIR) domain. Then, phoneme classification beyond the typically used techniques is extended towards exploring Deep Neural Networks (DNNs). This is done by combining Convolutional Neural Networks (CNNs) with audio data converted to the time-frequency space domain (i.e. spectrograms) and then exported as images. In this way a two-dimensional representation of speech feature space is employed. When preparing the phoneme dataset for CNNs, zero padding and interpolation techniques are used. The obtained results show an improvement in classification accuracy in the case of allophones of the phoneme /l/, when CNNs coupled with spectrogram representation are employed. Contrarily, in the case of vowel classification, the results are better for the approach based on pre-selected features and a conventional machine learning algorithm.