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 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).

    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 the fact that the
    upstream revision is not recorded in the artefact metadata.

    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 `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
    `tools/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.
