StyleTTS 2
An open-source text-to-speech model that runs locally, generates expressive speech and adapts to unfamiliar voices from reference audio.
StyleTTS 2 is a text-to-speech model for developers and speech researchers who want to generate expressive speech on their own hardware. It can choose a speaking style from the text without a reference recording.
- Expressive delivery: Generate speech with variation in pauses, speaking rate and emotion, or apply an emotional style to text that doesn't express that emotion itself.
- Voice adaptation: The LibriTTS model uses reference audio to reproduce voices it didn't encounter during training, including emotion and aspects of the recording's acoustic environment.
- Long-form narration: Synthesize longer passages as well as individual sentences, with pretrained LJSpeech and LibriTTS models available for single-speaker and multispeaker use.
- Custom voice training: Train or fine-tune models on your own speech data, including other languages when you use a suitable PL-BERT model; a multilingual PL-BERT is available.
The open-source Python code uses PyTorch and carries the MIT license. Inference runs locally with downloaded model checkpoints. The supplied LJSpeech and LibriTTS models cover English speech; other languages require language-appropriate training data and PL-BERT support. Training and fine-tuning can use NVIDIA GPUs, and fine-tuning supports a single GPU.
Logotyper
