Fine Tuning and Comparing Tacotron 2, Deep Voice 3, and FastSpeech 2 TTS Models in a Low Resource Environment

Abstract

Text-to-speech (TTS) models are used to generate speech from a sequence of characters provided as input. Existing TTS systems require a high-quality large dataset and vast computational resources for training. However, most of the publicly available datasets do not meet such standards, and access to powerful GPUs may not always be possible. Hence, in our work, we have successfully trained and compared TTS models, specifically Tacotron 2, FastSpeech 2, and Deep Voice 3 on a Tesla T4 GPU using a subset of the LJSpeechl.1 dataset. Subsequently, we have surveyed to analyze the performance of the models when trained on small datasets, and we discovered that the Tacotron 2 TTS model synthesized the most realistic sounding speeches. The survey revealed that the Tacotron 2 TTS model achieved a mean opinion score (MOS) at a 95% confidence interval of 4.25± 0.17, and sounded the most natural to our listeners when compared to the ground truth.