StyleTTS2: open source TTS with style diffusion outperforms Voicebox on intelligibility

StyleTTS2 is an open source speech synthesis system developed at Columbia University that produces voices so natural they are often indistinguishable from human in subjective tests. Its central idea is to treat vocal style (tone, rhythm, emotion) as a continuous vector and use diffusion to sample different styles in a controlled way. Thanks to adversarial training, the model learns to generate convincing audio even on subtle details like the micro-prosodic variations typical of human speech. It is fully open source (Apache 2.0) and has democratized access to professional-quality TTS for developers and researchers.