The paper explores the development of Automatic Speech Recognition (ASR) models for Armenian, by using data from two standard dialects (Eastern Armenian and Western Armenian). The goal is to develop a joint bi-variational model. We achieve state-of-the-art results. Results from our ASR experiments demonstrate the impact of dataset selection and data volume on model performance. The study reveals limited transferability between dialects, although integrating datasets from both dialects enhances overall performance. The paper underscores the importance of dataset diversity and volume in ASR model training for under-resourced languages like Armenian.

The Armenian language has many dialects that differ from each other syntactically, morphologically, and phonetically. In this work, we implement and evaluate models that determine the dialect of a given passage of text. The proposed models are evaluated for the three major variations of the Armenian language: Eastern, Western, and Classical. Previously, there were no instruments of dialect identification in the Armenian language. The paper presents three approaches: a statistical which relies on a stop words dictionary, a modified statistical one with a dictionary of most frequently encountered words, and the third one that is based on Facebook’s fastText language identification neural network model. Two types of neural network models were trained, one with the usage of pre-trained word embeddings and the other without. Approaches were tested on sentence-level and document-level data. The results show that the neural network-based method works sufficiently better than the statistical ones, achieving almost 98% accuracy at the sentence level and nearly 100% at the document level.