Development of method for using a neural network for voice identification taking into account specific accents

Abstract

The study aimed to investigate neural networks for voice identification with accented speech. The main models of Bidirectional Encoder Representations from Transformers (BERT) for Russian, English, Spanish, and Kazakh and their indicators were studied using the comparative and contrastive method. The main stages of creating a model for accent recognition include audio data preprocessing (noise removal, volume normalisation, fragmentation), extraction of low-frequency cepstral coefficients for the audio format suitable for analysis, Mel filtering, transformation of low-frequency cepstral coefficients for the model, training and evaluation. BERT models show different performances depending on the language. Language features such as morphology and syntax require unique customisation. For instance, BERT for Russian and Spanish incorporates declensions, and for Chinese – ambiguity and characters. The BERT for English reaches 90-96% accuracy, as the model was initially trained on English-language data. Multilingual BERT processes several languages, but the accuracy (70-85%) and F1-measure (70-80%) are lower than those of models configured for specific languages. The kazakhBERTmulti model demonstrates high accuracy (F1-measure – 0.68), outperforming Multilingual BERT Russian, and is better adapted to the Kazakh language with its agglutinative structure. The transformation of the low-frequency cepstral coefficients and using BERT achieved an accent recognition accuracy of 92%. More Kazakh data could have improved the model's accuracy, as accents are reflected in pronunciation, not spelling, so eliminating the spelling check focuses the model on accent.