Tortoise TTS V2

Tortoise TTS V2 is an open-source text-to-speech (TTS) program developed by James Betker, offering outstanding multi-voice capabilities and realistic prosody and intonation. It is equipped with various features such as random voice generation, using user-provided conditioning latents, and using pretrained models.

This model is characterized by utilizing both an autoregressive decoder and a diffusion decoder. This approach produces highly sophisticated results, but has the drawback of slow generation speed. For example, in a K80 GPU environment, generating a medium-sized sentence can take several minutes.

The core technology, the autoregressive decoder, works by predicting the next sound based on previous output values. This sequential dependency captures the rhythm, tone, and subtle nuances of speech, realizing more natural and human-like voice.