ITAcotron 2: Transfering English Speech Synthesis Architectures and Speech Features to Italian

End-to-end deep learning models have pushed forward significantly many tasks of Natural Language Processing (NLP). However, most of these models are trained for languages providing many resources (such as English), and their behaviour is hardly studied in other languages due to resource shortage. To cope with these situations, it is common practice to employ transfer learning. With this work, we wanted to explore the cross-language transferability of a Text-to-Speech (TTS) architecture and the re-usability of the surrounding components that complete a speech synthesis pipeline. To do so, we fine-tuned an English version of the Tacotron 2 TTS, with speaker conditioning, to Italian (hence ITAcotron 2). The human evaluation –carried on 70 subjects– showed that the language adaptation was indeed successful.