Speech & Audio
A text-to-speech model trained on 16 languages reached the best published results in most of those languages, its creators report

Most text-to-speech systems that imitate a new speaker without extra training work in only one language. In a 2024 paper at Interspeech (USB), researchers describe XTTS, a text-to-speech model trained on 16 languages. They report that it reached the best published results in most of those languages, and they released the model publicly.
The team built XTTS on an earlier model called Tortoise. They added modifications to support multilingual training, improve voice cloning, and speed up training and inference. Inference is the step where the model turns text into speech.
How does XTTS differ from earlier multilingual systems?
Earlier multilingual systems covered only a few high-resource or medium-resource languages. XTTS was trained on 16 languages, which widens the language coverage described in the paper.
Did it perform best in every language?
No. The authors report the best published results in most of the 16 languages, not all of them.
Is the system available to others?
Yes. The paper says the team made XTTS publicly available.
Limits and what's next
The abstract describes a model trained on 16 languages, but it does not name those languages or the evaluation data. The result covers most of those languages, not all of them. This is one study, and the abstract reports no independent replication by another team. The abstract also does not say how the model performed in the languages where it did not rank first.
Source: Edresson Casanova et al., “XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model,” Interspeech (USB), 2024.