Conditioning Text-to-Speech synthesis on dialect accent: a case study

Modern text-to-speech systems are modular in many different ways. In recent years, end-users gained the ability to control speech attributes such as degree of emotion, rhythm and timbre, along with other suprasegmental features. More ambitious objectives are related to modelling a combination of speakers and languages, e.g. to enable cross-speaker language transfer. Though, no prior work has been done on the more fine-grained analysis of regional accents. To fill this gap, in this thesis we present practical end-to-end solutions to synthesise speech while controlling within-country variations of the same language, and we do so for 6 different dialects of the British Isles. In particular, we first conduct an extensive study of the speaker verification field and tweak state-of-the-art embedding models to work with dialect accents. Then, we adapt standard acoustic models and voice conversion systems by conditioning them on dialect accent representations and finally compare our custom pipelines with a cutting-edge end-to-end architecture from the multi-lingual world. Results show that the adopted models are suitable and have enough capacity to accomplish the task of regional accent conversion. Indeed, we are able to produce speech closely resembling the selected speaker and dialect accent, where the most accurate synthesis is obtained via careful fine-tuning of the multi-lingual model to the multi-dialect case. Finally, we delineate limitations of our multi-stage approach and propose practical mitigations, to be explored in future work.