e2-tts-pytorch

Why we included this project

Voice assistants, audiobook pipelines, and dubbing workflows all run into the same wall: a TTS model that captures a new speaker's voice without retraining. This repo is a compact PyTorch implementation of E2-TTS, which is built for exactly that kind of zero-shot synthesis. It differs from the paper by using a multistream transformer for text and audio, with conditioning applied in every block, and it ships an optional trick that interpolates text to the audio length instead of predicting durations. Working end-to-end training code came from a community contributor, so the project holds up as a starting point rather than a paper reproduction. Lucas Newman's companion Nanospeech repo adds training recipes and an MLX port for Apple Silicon. It's research-grade software, so plan to handle preprocessing and dataset logistics yourself, but the codebase is small enough to reshape for your own pipeline.