Effective and Traceable Speech Synthesis for Low-Resource Languages
Supervisor: Mirko Marras
Co-Supervisor: Marco Manolo Manca
Candidate: Gabriele Carta
Abstract
Recent advances in speech synthesis have led to highly expressive and intelligible text-to-speech systems. However, these developments have predominantly targeted high-resource languages, leaving low-resource languages underrepresented and further widening the digital divide. In this thesis, we address this disparity by focusing on Sardinian, a language spoken on the island of Sardinia, characterized by limited digital resources and significant dialectal variation. We present a reproducible pipeline for speech data processing, including cleaning, alignment, and segmentation, and fine-tune two state-of-the-art multilingual text-to-speech models (XTTS-v2 and F5TTS) on two distinct Sardinian corpora. The resulting models are evaluated using established objective metrics to assess synthesis quality across dialectal varieties. Beyond synthesis performance, we examine the ethical implications associated with the uncontrolled use of generative speech technologies, particularly the risks of voice impersonation and misinformation. To address these concerns, we integrate an acoustic watermarking mechanism into the generated speech, enabling verifiable traceability and reliable differentiation between synthetic and human speech. Our findings contribute a principled and generalizable framework for responsible speech synthesis in low-resource contexts, with implications for both technological development and ethical deployment.
Audio Samples
Mannigos
De arrumbare... apo irbettau [...] dòighi oras, chi maridu meu achiat deghe, dòighi, [...] peri sèighi oras de fonderia e quindi apo irvetau... Nuorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Depimus nàrrere gasi, non l’amus a connòschere mai totu sa Sardigna, sa vera Sardigna, su veru sardu, ca b’at tantas cosas... Nuorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Mancu acatada [...] a cuìndixi annus non mindi podia mancu rendi contu e non mindi importàt nemmancu nudda, immoi invece medas cosas ddas seu cumprendendi immoi, quindi... Campidanese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Noi mesis a Taranto e poi cìncui annus in Sardegna, poi seus torraus a La Spezia, e deu sinceramenti non ci olia prus torrai a andai in Sardegna, deu dd’amu sa Sardegna però... Campidanese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Bi fit Lorenzo, devo nàrrere chi sempre l’atacaiat roba de su gènere, issu non aiat ispàtziu e tando mi so postu eo! Logudorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Apidu praticamente una agricoltura, ditzimus, arretrada, fossi su tipu de agricoltura chi t’ammentaisti sende pitzianna in Sardigna… Tzertu! Logudorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
SardinianVox
Volent organizzare. Poi si... sai sas iscolas riflettiant semper su.. s'andamentu 'e sa bidda sos pastores si diminùint... Nuorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Però afestaimus puru sa 'e Mesaustu, chi est sa die chi sos theracos cambiant mere o cunfirmaint o cambiaint, chi... Nuorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Connotu sa miniera poita acanta 'e su cungiau nostu pedra de fogu ddoi fut una miniera pròpiu cunfinanti, e deu andai a... Campidanese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Poita ca no ddu boint mancu intendi su paragoni, no si nd'importat de nudda. E puru nosu funt totu cosas chi eus passau e eus crèsciuì... Campidanese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
Poi... in s'iscola. E poi crachi die 'e festa, a su tempus de su fascismu, faiant unu pranzu, faiant e... Logudorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked
De pagu tale Ià, zente de pagu tale, cherzo narrere, tantu po fae riere Ià, e narada unu chi, nan chi si cheria torrare a coiuvare, e ca... Logudorese
Original
XTTS-v2
F5TTS
XTTS-v2 - Watermarked
F5TTS - Watermarked