DEVELOPMENT OF A MULTISPEAKER ARMENIAN TEXT-TO-SPEECH SYSTEM USING ARMUS CORPUS

Abstract

Text-to-speech synthesis for under-resourced languages faces significant challeng-es due to limited available datasets and lack of language-specific implementations. This paper presents the development and evaluation of a high-quality multispeaker Text-to-speech system for Armenian language based on VITS (Variational Inference with adversar-ial learning for end-to-end Text-to-speech) architecture. The system was trained on the ARMUS corpus, a comprehensive Armenian speech dataset containing approximately 21 hours of professional recordings from two male speakers. The VITS architecture combines conditional variational autoencoders with adversarial training to achieve end-to-end speech synthesis while preserving speaker identity and prosodic characteristics specific to Armeni-an phonology. The model implements specialized components including duration predic-tion for accurate phoneme timing, flow-based decoder for prosody modeling, posterior en-coder for latent representation, and waveform generator with hierarchical discriminators. Training was conducted over 1.5 million iterations using multiple loss functions including mel-spectrogram reconstruction, Kullback-Leibler divergence, adversarial losses, feature matching, and duration prediction. Comprehensive evaluation was performed using five objective metrics on 500 test utterances. Results demonstrate exceptional speaker similarity preservation with a score of 0.9366, good perceptual quality with PESQ score of 2.87, satis-factory intelligibility with STOI score of 0.7248, and accurate intonation reproduction with F0 correlation of 0.8834 and F0-RMSE of 8.40 Hz. Speaker verification analysis yielded an Equal Error Rate of 10%, confirming effective speaker distinction capability. The system successfully generates natural-sounding Armenian speech while maintaining distinct speak-er characteristics, representing a significant advancement in Armenian speech synthesis technology.