VALL-E
lastmod 2023-01-10created 2023-01-101 edits
VALL-E - a language model for speech synthesis made by Microsoft
Features
- Transformer-based Text-to-Speech model
- Can synthesize speech in any voice from just 3 seconds of that voice
- Far more natural and closer to the speaker than recent Zero-shot TTS models, and also preserves the speaker's emotion and acoustic environment
- The old pipeline was phoneme → mel-spectrogram → waveform,
- VALL-E is phoneme → discrete code → waveform
- Can be combined with various speech synthesis applications and AI models such as GPT-3