zerre-tts ========= This product includes software derived from tiny-voca (https://github.com/n4ze3m/tiny-voca), Copyright Muhammed Nazeem, licensed under the Apache License, Version 2.0. The model architecture, the training recipe and parts of the source tree in src/zerre/ (model.py, losses.py, data.py, train.py, export.py, align.py, annotate.py, prepare.py, evaluate.py, journal.py, inspect.py, refine.py) come from tiny-voca. They were modified for this project: the English phoneme front end was replaced by a language-pack layer, the token inventory and normalisation are new, and platform-specific code (file locking, executable discovery, paths) was made portable. The Turkish dataset, annotations, checkpoints and the language packs added here are original to this project. Third-party components used at runtime or during data preparation: * eSpeak NG (GPL-3.0) — external program used only to turn text into phonemes. It is not linked into the model and is not distributed with these weights. * PyTorch (BSD-3-Clause), NumPy (BSD-3-Clause), soundfile (BSD-3-Clause), filelock (Unlicense). * For alignment and evaluation, optionally: torchaudio, transformers (Apache-2.0) and faster-whisper (MIT). Their weights are not part of the released model.