Enhancing TTS Performance for Noisy and Low-Resource Languages Using Advanced Noise Removal Techniques

This study investigates the impact of advanced noise removal techniques on Text-to-Speech (TTS) performance for low-resource languages, focusing on Armenian as a case study. Text-to-speech systems for low-resource languages often perform poorly due to limited and noisy datasets. The hypothesis is that applying state-of-the-art noise removal methods can significantly enhance TTS quality in these challenging scenarios. Using the Armenian subset of the Common Voice dataset, a baseline was first established by training a Tacotron2 TTS model on the original noisy data. SpeechBrain, an advanced speech processing toolkit, was then applied to clean the dataset. A second Tacotron2 model was trained on this noise-reduced dataset. Both models were evaluated using subjective Mean Opinion Score (MOS) tests. The results demonstrated an improvement in TTS quality for the model trained on the noise-removed dataset, with a 10.3% increase in MOS compared to the baseline. This study highlights the potential of noise removal techniques, specifically SpeechBrain, to boost TTS performance in low-resource scenarios, paving the way for improved speech synthesis in a wider range of languages. The findings have important implications for developing more inclusive and higher-quality TTS systems for underrepresented languages.