Voxtral 3

Voxtral 3 is a text-to-speech (TTS) model developed by researchers at the University of California, Berkeley. The model uses a transformer-based architecture and is trained on a dataset of 1,000 hours of speech data, which is significantly larger than the 50 hours used to train the previous VoxTract model. This expanded training dataset allows Voxtral 3 to produce more natural and expressive speech. The researchers claim that Voxtral 3's speech is indistinguishable from that of a human, with improvements in both naturalness and intelligibility. In particular, Voxtral 3 is said to be able to produce speech that is more natural and less robotic than previous TTS models. The researchers also mention that Voxtral 3 is able to learn from a wide range of voices and accents, including those from different regions and languages.