// the take
A from-scratch implementation of CycleGAN for non-parallel voice conversion, converting between two speakers' MCEP features without needing paired training data. Aimed at ML researchers or students who want to see the CycleGAN-VC paper translated into working TensorFlow code rather than at anyone building a production voice tool.
Clean separation of preprocessing, model, and training/conversion scripts makes it easy to follow the paper-to-code mapping. Ships pretrained weights plus actual converted .wav demo files, including a side-by-side against the original NTT authors' output, so you can judge quality without training anything yourself. The Dockerfile pins the exact TF 1.8/CUDA environment, which matters a lot here given how fragile that stack is to reproduce manually.
Hard-pinned to Python 3.5 and TensorFlow 1.8 — both long dead — so getting this running on any current GPU driver/CUDA setup means fighting compatibility issues the repo never addresses. The repo's own to-do list admits there are no evaluation metrics; quality assessment is literally 'using our bare ear,' which is fine for a demo but useless for comparing model variants objectively. Each trained model handles exactly one speaker pair (A2B/B2A), so multi-speaker conversion means retraining from scratch per pair, and training takes 1000+ epochs — expensive even on the TITAN X the author used, worse on anything without a comparable GPU.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →