Machine Learning Algorithms Based on Hidden Markov Models in Low-Speed Speech Codecs for Assessing Speech Quality

A machine learning algorithm for a low-speed primary coding algorithm based on automatic recognition of speech signal fragments (SS) using hidden Markov models (HMM) has been developed. When assessing the quality of SS at transmitter output, spectral dynamics algorithm is used, as well as complex algorithms for assessing speech quality. Reduction of redundancy in SS is provided by using speech recognition and synthesis algorithms. Training with HMM to solve speech recognition problem included the definition of such parameters as: matrix of probabilities of transitions between different states of the model. To solve these problems two iterative algorithms were used: forward-backward algorithm and Baum-Welsh algorithm. As an experiment, using Hidden Markov Model Toolkit (HTK) software model, an independent monophonic speech model of Russian language was built. The model was based on the recognition from a list of phonemes consisting of 52 phonetic units and Russian language dictionary including 500000 words. Well-known algorithms to assess speech quality at low-speed codec output are investigated in the interests of machine learning. It is shown that the use of these algorithms does not provide necessary correlation of objective and subjective algorithms to assess speech quality under the action of acoustic interference. Algorithms for objective quality assessment that ensure the correlation of objective and subjective assessments of speech quality with the accuracy of 0.5 points in accordance with GOST R 50840-95 are proposed. The research has shown that the use of these algorithms in the encoder allows increasing speech quality at the output of low-speed codec with its machine learning by 0.5 ... 1 point in accordance with GOST R 50840-95.