Javascript must be enabled for the correct page display

Synthesising Proto-Indo-European using Phonological Features for Zero-Shot Synthesis

Ivnova, Victoria (2023) Synthesising Proto-Indo-European using Phonological Features for Zero-Shot Synthesis. Master thesis, Voice Technology (VT).

[img]
Preview
PDF
MA 53226 V Ivanova.pdf

Download (1MB) | Preview

Abstract

Proto-Indo-European is a reconstructed language, from which the biggest language family, Indo-European, evolved. Linguists have reconstructed its phonology through the comparative method, analysing cognate words in its daughter languages. Some attempts at automation this process have been made, but fewer have attempted to take it a step further and synthesise its sound. This task could be seen as a zero-shot synthesis problem, meaning that it needs to be synthesised without any training data for a model to learn from. We ask whether it is possible for this task to be achieved through the means of zero-shot synthesis using phonological features as input. Models utilizing this technique have been shown to produce successfully unseen languages in code-switching tasks and even synthesising unseen phonemes. We opt to use the IMS-Toucan toolkit, which is mostly build upon the FastSpeech2 architecture, with some additions, such as the use of the LAML optimizing framework. The toolkit is modular and we can modify its text-processing and phonemization modules to handle Proto-Indo-European input. Further, we fine-tune the multilingual model on Abkhaz, which has some similar features to Proto-Indo-European. Our results find that our method improves significantly the naturalness of the synthesised speech in comparison to previous attempts at synthesising Proto-Indo-European, but the fine-tuning yields no significant improvement over the pre-trained model. The user-friendly web app that we built is a useful tool for education or entertainment purposes. What is more, we believe that our system could be beneficial to language revitalization tasks and combined with other methods for automation of the reconstruction process, it could lead to better success in the efforts of keeping the languages of the world alive.

Item Type: Thesis (Master)
Name supervisor: Coler, M.L. and Do, T.P.
Date Deposited: 12 Sep 2023 11:10
Last Modified: 12 Sep 2023 11:10
URI: https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/371

Actions (login required)

View Item View Item