Researchers have released CVSS-X, a synthetic speech-to-speech translation corpus covering English input and 28 target languages from 12 language families. It contains roughly 240,000 parallel speech pairs per language and more than 16,000 hours in total.
The dataset reverses and expands the direction covered by CVSS, which translates speech from 21 languages into English. Used together, the two collections can support experiments in bidirectional and multilingual spoken translation rather than an English-only destination.
CVSS-X comes in two generated variants. CVSS-X-C uses two canonical voices for each language, while CVSS-X-T applies cross-lingual voice cloning. The authors report translation quality comparable to CVSS and consistent results across typologically different languages.
All of the speech is synthetic, so the corpus does not reproduce the full range of real speakers, accents, recording conditions and conversational behavior. The dataset is available under a CC-BY-NC 4.0 license, which restricts commercial use, and the code is public. The paper is a new arXiv preprint; downstream systems will still need evaluation on natural speech and by speakers of the target languages.