Antalia-2 Mini
Copyright 2026 PatientDesk AI

The code, the text normaliser and the model weights are licensed under the Apache License,
Version 2.0 (see LICENSE).

This package includes PatientDesk AI's Turkish text normaliser (antalia_mini/text/normalize.py,
antalia_mini/text/foreign.py, antalia_mini/text/data/lexicon.json, antalia_mini/text/data/tr_vocab.json),
released under the same licence. Vendored verbatim from turkish-tts commit 3ed2e60 (2026-10-07).

At run time the normaliser uses the CMU Pronouncing Dictionary through the `cmudict` package
(Copyright Carnegie Mellon University, BSD 2-Clause licence); it is a dependency, not included here.

The model weights were trained on synthetic speech generated with VoxCPM2 (OpenBMB, Apache-2.0).

The model reimplements published methods: monotonic alignment search (Glow-TTS, Kim et al. 2020;
Matcha-TTS, Mehta et al. 2024), flow matching (Lipman et al. 2023), shortcut models (Frans et al. 2024),
adaLN-single (PixArt-alpha, Chen et al. 2023), ConvNeXt (Liu et al. 2022), Vocos (Siuzdak 2023),
HiFi-GAN's multi-period discriminator (Kong et al. 2020), UnivNet's multi-resolution discriminator
(Jang et al. 2021), RoPE (Su et al. 2021), SwiGLU (Shazeer 2020). No code from those projects is included.
