MegaTTS3

ByteDance released MegaTTS3 in March 2025 as a zero-shot text to speech system built around a sparse-alignment latent diffusion transformer whose TTS backbone carries only 0.45B parameters. It speaks Chinese and English with code-switching, and exposes an accent intensity control that preserves or neutralizes an accent. The repository ships a command line inference script, a Gradio web API, and modules for speech-text alignment, grapheme-to-phoneme conversion, and the WavVAE codec that compresses 24 kHz speech into 25 Hz latents, with weights on Hugging Face under ByteDance/MegaTTS3. One limitation matters: for safety reasons ByteDance withholds the WavVAE encoder weights, so cloning a new speaker means submitting a clip under 24 seconds to the voice request queue and waiting for the matching .npy latent. Setup uses Python 3.10 and PyTorch 2.6, with Linux, Windows, and Docker paths documented, and CPU inference takes roughly 30 seconds for 10 diffusion steps, so a CUDA GPU is the practical choice. The code is Apache-2.0 though the authors state the project is primarily intended for academic purposes. The repo has passed 6,000 stars.