Real human voices.

VoxGuess currently uses a curated selection from the Google FLEURS-R speech dataset: 10,000 recordings spanning 100 languages.

Source and licence

Google, FLEURS-R: A Restored Multilingual Speech Corpus for Generation Tasks. The dataset is published under the Creative Commons Attribution 4.0 International licence.

Processing

FLEURS-R uses Google’s Miipher model to restore human speech recordings. VoxGuess trims leading silence and normalizes loudness for consistent playback. VoxGuess does not alter the language being spoken. Recordings that appear silent, distorted or unsuitable are removed through automated checks and listening review.

Report a clip

If a recording sounds wrong or contains unexpected noise, please send us the language, round and approximate time.