loudkit Copyright 2026 LoudReader This product includes software developed at LoudReader. Licensed under the Apache License, Version 2.0. See LICENSE. -------------------------------------------------------------------------------- Third-party notices -------------------------------------------------------------------------------- Chatterbox TTS Copyright (c) 2025 Resemble AI Licensed under the MIT License. https://github.com/resemble-ai/chatterbox The weights distributed with this project — the `loudr` series — are derived from Chatterbox and optimised for real-time synthesis. Parts of the pipeline were retrained; others, among them the tokenizer and speaker encoder that enrollment uses, are carried from the upstream artefact unchanged and are redistributed here under the MIT terms above. The architectures in `python/loudkit/models/` are reimplementations written against those weights; they are not copies of the upstream source, but the weights they load could not exist without it. Voice profiles The shipped voice profiles — twenty-eight enrollments across ten languages, distributed on the Hugging Face repository under voices/profiles/ — are enrolled from the recordings catalogued in `docs/voices/roster/provenance.json`, which records donor or speaker, source, licence, consent basis, reference construction and sha256 for every profile, reference and sample. None of them are Chatterbox artefacts. Every source below was recorded or released for speech- technology use, and each profile ships under its own terms: OHF-Voice / NabuCasa voice-datasets — CC0-1.0 https://github.com/NabuCasa/voice-datasets Voice donations recorded expressly for building TTS voices: darkman, gosia (pl); joe, kathleen (en); kerstin (de); pim, nathalie (nl); dave (es); paola (it); tugao (pt). The kathleen reference reads prompts from the CMU ARCTIC prompt list; those prompt texts come from public-domain novels and impose no restriction on the recording itself. Thorsten-Voice, TV-44kHz-Full — CC0-1.0 https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full thorsten (de), under the donor's own public-domain dedication. NST Swedish and NST Danish speech databases — CC0-1.0 Språkbanken, National Library of Norway: https://www.nb.no/sprakbanken/en/resource-catalogue/oai-nb-no-sbr-17/ https://huggingface.co/datasets/alexandrainst/nst-da nils, selma (sv); soren, freja (da). Kyutai tts-voices — CC-BY-4.0 https://huggingface.co/kyutai/tts-voices henri, colette (fr). Kyutai Voice Donations — CC0-1.0 https://huggingface.co/kyutai/tts-voices, under voice-donations/ clara, emma, henry, lucy, miles, oliver, oscar, sophie (en). Donations to the Unmute Voice Donation Project, whose terms expressly cover speech synthesis, voice embeddings and commercial use: https://kyutai.org/legal/Terms%20of%20Use%20-%20Unmute%20Voice%20Donation%20Project%20v1.pdf The names above are product aliases, not donor identities; the roster records each donation's speaker id, revision and sha256. CML-TTS — CC-BY-4.0 https://huggingface.co/datasets/ylacombe/cml-tts carmen (es), ines (pt). LibriVox-derived recordings; the dataset's terms credit the original LibriVox readers. Multilingual LibriSpeech — CC-BY-4.0 https://huggingface.co/datasets/facebook/multilingual_librispeech dante (it). LibriVox-derived; likewise credits the readers. Voice encoder (`ve.safetensors`) The 5.7 MB utterance voice encoder that drives voice cloning is an export of the `VoiceEncoder` model from the Chatterbox T3 architecture (Resemble AI, MIT — see above), produced by this project's own enrollment export (this project's enrollment export tooling). It is a derived work of the Chatterbox weights and is redistributed under the same MIT terms. `docs/PROVENANCE-voice-encoder.md` records the artefact's size and sha256, its tensors, the upstream repository and licence, and an exact tensor match with ResembleAI/chatterbox revision 5bb1f6ee58e50c3b8d408bc82a6d3740c2db6e18. The original export's checkout revision was not recorded. MIT License text: Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. Enrollment architectures `python/loudkit/models/enroll.py` reimplements four inference paths whose designs come from the projects below. The weights those modules load are part of the Chatterbox-derived checkpoint (MIT, above); what is owed here is attribution for the architectures, all of them permissively licensed and all compatible with this project's Apache-2.0 terms. CosyVoice — Apache License 2.0 Copyright FunAudioLLM / Alibaba Group https://github.com/FunAudioLLM/CosyVoice The S3 speech tokenizer's shape (whisper-style encoder, FSQ quantiser, 25 Hz token rate) and the matcha-recipe mel the flow conditions on. S3Tokenizer — Apache License 2.0 Copyright Xingchen Song https://github.com/xingchensong/S3Tokenizer The standalone tokenizer extraction this port follows, including the FSMN memory over the attention values. 3D-Speaker (CAM++) — Apache License 2.0 Copyright ModelScope / Alibaba Group https://github.com/modelscope/3D-Speaker The CAM++ x-vector encoder: FCM front end, TDNN-dense body with context-aware masking, statistics pooling. FunASR — MIT License Copyright ModelScope / Alibaba Group https://github.com/modelscope/FunASR The FSMN formulation the tokenizer's attention block carries. Real-Time-Voice-Cloning — MIT License Modified & original work Copyright (c) 2019 Corentin Jemine Original work Copyright (c) 2018 Rayhane Mama Original work Copyright (c) 2019 fatchord Original work Copyright (c) 2015 braindead https://github.com/CorentinJ/Real-Time-Voice-Cloning The utterance voice encoder: 3-layer LSTM over 40-mel partials, L2-normalised, which is the recipe its weights were trained against. CMUdict (Carnegie Mellon Pronouncing Dictionary) Copyright (C) 1993-2015 Carnegie Mellon University. All rights reserved. `python/loudkit/models/data/pl_en_respell.json` is generated from CMUdict by `tools/gen_pl_respell.py`, which maps ARPAbet phonemes to Polish orthography. It is a derived work and it ships in the distribution, so the notice below is reproduced as clause 2 requires. Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. The contents of this file are deemed to be source code. 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. This work was supported in part by funding from the Defense Advanced Research Projects Agency, the Office of Naval Research and the National Science Foundation of the United States of America, and by member companies of the Carnegie Mellon Sphinx Speech Consortium. We acknowledge the contributions of many volunteers to the expansion and improvement of this dictionary. THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. Unicode CLDR Copyright © 1991-2026 Unicode, Inc. All rights reserved. Licensed under the Unicode License v3 — https://www.unicode.org/license.txt `tests/data/cldr/*.ssv` are CLDR's own RBNF spellout test data, imported verbatim with their copyright header intact, and `tests/data/conformance/numbers_cldr.json` is the subset mapped onto rulesets this project implements. Test corpus only: none of it is read at runtime and none of it ships in the wheel or the sdist. It is a differential test corpus, independent of this implementation. NST pronunciation lexicons The Swedish, Danish and Norwegian NST lexicons that `research/fetch_nst_lexicons.py` downloads are public domain (CC0), from Språkbanken Tal. They are fetched on demand and never vendored, so nothing of theirs is redistributed here. Philox-4x32-10 The counter-based random number generator in `python/loudkit/rng.py` implements the algorithm described in: J. K. Salmon, M. A. Moraes, R. O. Dror, D. E. Shaw. "Parallel Random Numbers: As Easy as 1, 2, 3." SC11, 2011. The known-answer test vectors are from the Random123 reference library (D. E. Shaw Research), used for verification only. DSP filterbank and window tables `python/loudkit/models/data/dsp/*.f32`, and the copies each port carries (`swift/LoudKit/Resources/*.f32`, `rust/src/enroll_data/*.f32`, `go/enroll/data/*.f32`, `js/src/dspData.ts`), are numeric tables generated by `tools/gen_dsp_assets.py`. They are precomputed rather than derived at runtime so that five implementations multiply exactly the same numbers; the values come from four upstreams: librosa (ISC) — `librosa.filters.mel`, the Slaney-style mel filterbanks for the mel decoder (24 kHz / 1920-point) and the utterance voice encoder (16 kHz / 400-point). PyTorch and SciPy (BSD-3-Clause) — `torch.hann_window` and `scipy.signal.get_window`, the analysis windows, including the Povey variant (a Hann window raised to 0.85). Kaldi, via torchaudio's `torchaudio.compliance.kaldi` (Apache-2.0) — `get_mel_banks`, the speaker-encoder filterbank. Chatterbox (MIT) — the S3 tokenizer's mel filterbank, read out of the upstream checkpoint rather than recomputed, so the tokenizer sees the filters it was trained with. Every one of these licences is compatible with Apache-2.0; Kaldi and librosa ask that attribution travel with the values, which is what this section is. `python/loudkit/models/data/dsp/manifest.json` records the same provenance per file.