Hangry LabsUseful tools, easy to run

Third-Party Notices

Melo TTS combines software, data, model weights, and browser assets from multiple projects. The combined application source is distributed under AGPL-3.0-only, but that does not replace the licenses or notices that apply to individual components.

This inventory was reviewed on 2026-10-07. Links identify the upstream source and model cards used for the review. Redistributors should repeat the review before changing a dependency or model revision.

Source provenance

Hangry Labs modifications

Docker packaging, the browser application, API integration, documentation, release tooling, and other fork-specific modifications are copyright 2026 Hangry Labs and contributors and are distributed under AGPL-3.0-only.

Bert-VITS2

The inference architecture and portions of the text-processing implementation are derived from Bert-VITS2, distributed under GNU AGPL version 3. This inherited strong-copyleft code is the reason the combined Melo TTS application is distributed under AGPL-3.0-only. The complete AGPL text is in LICENSE.

Original MeloTTS

This project is based on MeloTTS by Wenliang Zhao, Xumin Yu, Zengyi Qin, MyShell.ai, and its contributors. MyShell.ai published the original MeloTTS repository with an MIT license and the notice:

Copyright (c) 2024 MyShell.ai

The preserved license is in LICENSES/MIT-MeloTTS.txt. The combined project is nevertheless distributed under AGPL-3.0-only because it also contains AGPL-covered Bert-VITS2-derived material.

PaddleSpeech tone sandhi

melo/text/tone_sandhi.py retains the PaddlePaddle copyright and Apache-2.0 header from PaddleSpeech. The Apache License 2.0 is reproduced in LICENSES/Apache-2.0.txt.

Tacotron number normalization

melo/text/english_utils/number_norm.py contains code originating in keithito/tacotron, copyright 2017 Keith Ito and licensed under MIT. The preserved license is in LICENSES/MIT-Tacotron.txt.

CMU Pronouncing Dictionary

melo/text/cmudict.rep contains the CMU Pronouncing Dictionary 0.6. Its original copyright and unrestricted-use notice are retained at the beginning of that file.

Bundled browser components

  • WaveSurfer.js: BSD-3-Clause. Its license is retained in melo/standalone_ui/static/vendor/wavesurfer/LICENSE.
  • Lucide: ISC, with MIT-licensed Feather icons. Its notices are retained in melo/standalone_ui/static/vendor/lucide/LICENSE.

    Runtime dependencies

Python packages remain under their own licenses. Their package metadata and license files are preserved in each installed .dist-info directory in the Docker image. Notable copyleft or separately licensed runtime components include:

  • pykakasi: GPL-3.0-or-later.
  • distance: GPL-2.0-or-later, as declared by its distributed package license.
  • num2words: LGPL.
  • SciPy: BSD-3-Clause. Its license and bundled binary dependency notices are retained in its installed .dist-info directory.
  • FFmpeg and its Debian libraries: the Debian build declares GPL-3.0-or-later as its effective license. Its licenses and source-package notices are retained under /usr/share/doc in the image.
  • NVIDIA CUDA runtime libraries installed with the CUDA-enabled PyTorch wheels: NVIDIA proprietary terms, retained in their .dist-info/licenses directories.

The AGPL license for Melo TTS does not replace these component licenses.

Models and language resources

Melo TTS voice checkpoints

The standard image embeds the following MyShell.ai model repositories, whose model cards declare MIT:

  • myshell-ai/MeloTTS-English
  • myshell-ai/MeloTTS-English-v2
  • myshell-ai/MeloTTS-English-v3
  • myshell-ai/MeloTTS-French
  • myshell-ai/MeloTTS-Japanese
  • myshell-ai/MeloTTS-Chinese

The English-family image embeds only the three English repositories. The MIT terms reproduced in LICENSES/MIT-MeloTTS.txt apply to these MyShell.ai model artifacts according to their published model-card metadata.

Spanish myshell-ai/MeloTTS-Spanish and Korean myshell-ai/MeloTTS-Korean checkpoints are not embedded in published images. An operator may explicitly download either pinned checkpoint from Hugging Face into their persistent local volume after reviewing the associated encoder terms below.

Language encoders

  • google-bert/bert-base-uncased: Apache-2.0.
  • google-bert/bert-base-multilingual-uncased: Apache-2.0.
  • dbmdz/bert-base-french-europeana-cased: MIT.
  • tohoku-nlp/bert-base-japanese-v3: Apache-2.0.
  • dccuchile/bert-base-spanish-wwm-uncased (BETO) is not distributed in the Docker image. Its model card says CC BY 4.0 best describes the authors’ intentions and warns that some training datasets may have incompatible terms, particularly for commercial use. Users should review the BETO model card for their use case and provide attribution to the BETO authors.
  • kykim/bert-kor-base is not distributed in the Docker image. Its Hugging Face model card does not declare a license. The linked LMkor project is Apache-2.0, but its README instructs commercial model users to arrange a free MOU with the author. Hangry Labs does not grant rights to it; operators must explicitly review its terms before the application downloads it from Hugging Face into their own persistent volume.

Dictionaries and language data

  • UniDic is distributed under its upstream GPL/LGPL/BSD choice. The downloaded dictionary keeps its COPYING, BSD, GPL, LGPL, and AUTHORS files inside the installed unidic/dicdir/licenses directory.
  • unidic-lite retains the corresponding dictionary notices in its installed package.
  • NLTK resources and pronunciation data retain their upstream notices in the installed NLTK data/package directories.

Generated audio

The software license does not normally apply to audio generated from user input. Users remain responsible for rights in their input, selected model, voice use, and resulting audio under applicable law and the model-specific terms above.