Comparison of neural network architectures for recognizing Russian speech with a foreign accent

Abstract

Automatic speech recognition systems are actively developing thanks to the widespread use of deep learning methods. However, when working with Russian speech pronounced by speakers of other languages, most algorithms encounter a decrease in accuracy. Accent specificity affects the duration of sounds, articulatory transitions, and prosodic  characteristics, which makes it difficult to correctly identify acoustic features and subsequently transcribe them. The scientific literature describes many approaches to speech processing, including deep, convolutional, and recurrent models, as well as hybrid architectures that use attention mechanisms. Each of them responds differently to pronunciation variability and accent intensity. Taking these features into account, this work aims to study the behavior of various neural network architectures in recognizing Russian speech with a foreign accent and to analyze their results on a corpus of our own recordings. The evaluation is based on WER, CER, and Accuracy metrics, which allows us to identify models that demonstrate the greatest resistance to accent distortions and are capable of working with limited and heterogeneous data.