Transfer Learning for Text-to-speech Systems in Romanian Language

Abstract

Text-to-speech systems in Romanian are constantly improving, the main problem being speaker-dependent recognition. In other words, manually matching phonemes to waveforms for each record. Today, these systems are based on deep neural networks that require large amounts of data, which is expensive and difficult. With regard to voice cloning, the capabilities of speech synthesis systems are still limited. In the classical manner, the process consists of manually transcribing and aligning phonemes with each audio recording of the target speaker. This article presents a possible solution for developing a text-speech synthesis system, using the transfer learning method, which can be trained on corpora in Romanian small language. In fact, it is only about audio elements and their transcription, without the need to collect any other data. The results of voice cloning on these small datasets are promising, exceeding the accuracy of other deep learning architectures.