Phoneme Recognition Using Interlaced Derivative Pattern and Co-occurrence Matrix Method

Abstract

Accurate recognition of phonemes has always been a difficult point in speech recognition. This paper attempts to use the co-occurrence matrix method and interlaced derivative patterns method in image processing to complete the task of phoneme recognition. First, perform segmentation of speech based on the spectral energy and spectral centroid. Second, perform feature extraction, and calculate the four features of the co-occurrence matrix method, interlaced derivative pattern method, multi-dimensional voice program parameters and gammatone frequency cepstral coefficients of the phoneme segment. Among them, the co-occurrence matrix method and interlaced derivative pattern method are calculated based on the gammatone spectrum of the voice. Then, use random forest to calculate the importance of feature vectors, and select the top 30 features that random forest considers the most important as the features used in this article. Finally, random forest is used for classification. Note that the phoneme data sets used in this article are all downloaded from YouTube, and the classification we do is vowels, semivowels and consonants. To our best knowledge, this is the first paper that presents the phoneme recognition using interlaced derivative pattern and co-occurrence matrix method.