KimiTalk Speech Helper Notice
Version: 1.2
Date: 2026-05-26
The plain-text LocalSpeechHelper notice is available at https://kimikon.cc/kimitalk/speech-helper/NOTICE.txt.
================================================================================
KIMITALK LOCAL SPEECH HELPER — OPEN-SOURCE NOTICE & ACKNOWLEDGEMENTS
================================================================================
Component: LocalSpeechHelper.app (Speech Output Add-On for KimiTalk)
Version: See helper-metadata.json (helper_version)
Vendor: KIMIKON e.U., Vienna, Austria
Mirror: https://updates.kimikon.cc/kimitalk/speech-helper/<version>/NOTICE.txt
Date: 2026-05-26 — license hygiene pass
Copyright (c) 2024-2026 KIMIKON e.U., Vienna, Austria.
This Helper bundles open-source software and consumes local model caches prepared
by KimiTalk. In the ordinary release path, KimiTalk prepares those caches from
KIMIKON's `models.kimikon.cc` mirror, verifies SHA-256 checksums, then starts the
Helper in offline runtime mode. Upstream repository names remain documented
below for provenance and license attribution.
Full license texts live in the adjacent LICENSES/ directory.
This file applies ONLY to the LocalSpeechHelper.app DMG. The main KimiTalk.app
bundle has its own NOTICE / Acknowledgements view inside the app.
================================================================================
PART 1 — BUNDLED SOFTWARE IN LocalSpeechHelper.app
================================================================================
The Helper is a Python-based subprocess launched by KimiTalk. The following
components are bundled (statically copied) into the .app:
--------------------------------------------------------------------------------
1.1 Python Software Foundation Runtime PSF License v2
--------------------------------------------------------------------------------
Copyright (c) 2001-2026 Python Software Foundation. All Rights Reserved.
Python is licensed under the PSF License v2, with additional licenses for
bundled components (OpenSSL, zlib, libffi, expat). See LICENSES/PSF-Python.txt
for the full license text stack.
--------------------------------------------------------------------------------
1.2 PyTorch (torch) BSD 3-Clause
--------------------------------------------------------------------------------
Copyright (c) 2016-present, Meta Platforms, Inc. and affiliates.
Additional copyrights: Idiap Research Institute, DeepMind, NEC Labs, NYU,
Google, Kakao Brain, Cruise LLC, Tri Dao, Arm Limited, and PyTorch
contributors.
The bundled PyTorch wheel includes libiomp5 (Intel OpenMP runtime) under the
Apache License 2.0 with LLVM Exception.
See LICENSES/BSD-3-Clause.txt and LICENSES/Apache-2.0-FULL.txt.
--------------------------------------------------------------------------------
1.3 torchaudio BSD 2-Clause
--------------------------------------------------------------------------------
Copyright (c) 2017 Facebook, Inc.
See LICENSES/BSD-2-Clause.txt.
--------------------------------------------------------------------------------
1.4 Additional BSD/Apache Python dependencies
--------------------------------------------------------------------------------
- numpy — BSD 3-Clause
- scipy — BSD 3-Clause
- soundfile — BSD 3-Clause (Python wrapper)
- huggingface-hub — Apache License 2.0
- transformers — Apache License 2.0
- tokenizers — Apache License 2.0
- safetensors — Apache License 2.0
- mlx (Apple) — MIT License
- mlx-lm — MIT License
- mlx-audio — MIT License
- librosa — ISC License (compatible with BSD/Apache)
Full text references in LICENSES/Apache-2.0-FULL.txt, BSD-2-Clause.txt,
BSD-3-Clause.txt.
--------------------------------------------------------------------------------
1.5 libsndfile LGPL-2.1+
--------------------------------------------------------------------------------
Copyright (c) 1999-2024 Erik de Castro Lopo and the libsndfile contributors.
libsndfile is the native C library underlying the python-soundfile module. It
is dynamically loaded by the Python soundfile wheel via CFFI.
LICENSE: GNU Lesser General Public License, version 2.1 or later. See
LICENSES/LGPL-2.1-FULL.txt for the full license text.
SOURCE-OFFER (LGPL-2.1 § 6):
KIMIKON e.U. provides, upon written request to legal@kimikon.cc, the
unmodified libsndfile source corresponding to the bundled version, for
a fee not exceeding the cost of physical distribution. Please include
the LocalSpeechHelper version (visible in the helper-metadata.json /
KimiTalk About dialog) so we can ship the exact matching source.
--------------------------------------------------------------------------------
1.6 AudioSeal (Meta, facebookresearch/audioseal) MIT License
--------------------------------------------------------------------------------
Copyright (c) 2024 Meta Platforms, Inc. and affiliates.
STATUS: ACTIVE (Phase 19, Helper v0.6.0).
AudioSeal is bundled with this Helper as ONNX models in Resources/Models/
(generator + detector, 16-bit payload). The Helper invokes AudioSeal embed
and detect via the provenance.* NDJSON subcommands. C2PA belongs to this
Local Speech Helper: the Helper bundles c2pa-python and the native c2pa-rs
library shipped inside that wheel, and invokes them through the same
provenance channel. See sections 1.7 below, LICENSES/Apache-2.0-FULL.txt,
and LICENSES/c2pa-python-MIT.txt.
Phase 19 migrated the heavy provenance runtime (libc2pa_c.dylib and AudioSeal
ONNX models) out of the KimiTalk main app into this Helper, so the main DMG
stays small. Helper-backed AI-audio export requires this provenance runtime and
fails closed if mandatory AudioSeal/C2PA provenance cannot be created.
See LICENSES/AudioSeal-MIT.txt.
--------------------------------------------------------------------------------
1.7 C2PA / c2pa-python / c2pa-rs Apache 2.0 OR MIT
--------------------------------------------------------------------------------
c2pa-python is the official Python binding for the C2PA Content Authenticity
SDK. It wraps and bundles the native c2pa-rs library used for C2PA signing and
verification.
STATUS: ACTIVE (Helper provenance runtime).
KimiTalk.app does not link or bundle libc2pa. Helper-backed AI-audio export
uses this Helper process for C2PA manifests; export fails closed if mandatory
C2PA provenance cannot be created.
See LICENSES/Apache-2.0-FULL.txt and LICENSES/c2pa-python-MIT.txt.
================================================================================
PART 2 — RUNTIME-PREPARED MODEL WEIGHTS (KIMIKON/R2 mirror, not bundled in DMG)
================================================================================
Model weights are NOT bundled in this DMG. For release models, KimiTalk prepares
them into the local cache from the KIMIKON/R2 mirror at `models.kimikon.cc` and
verifies every manifest/file with SHA-256 before the Helper loads them. Licenses
and required notices are reproduced here because KIMIKON e.U. orchestrates model
preparation and documents the upstream model sources.
--------------------------------------------------------------------------------
2.1 hexgrad / Kokoro-82M Apache License 2.0
--------------------------------------------------------------------------------
Copyright (c) 2024 hexgrad and Kokoro contributors.
Source: https://huggingface.co/hexgrad/Kokoro-82M
https://github.com/hexgrad/kokoro
MLX: https://huggingface.co/mlx-community/Kokoro-82M-bf16
This product is not endorsed by or affiliated with hexgrad. "Kokoro" is the
project name of the upstream model; no trademark license is implied
(Apache 2.0 § 6).
--------------------------------------------------------------------------------
2.2 Qwen3-TTS-12Hz-1.7B-CustomVoice (Preset TTS, 9 voices) Apache License 2.0
2.3 Qwen3-TTS-12Hz-1.7B-Base (Own-Voice Cloning model) Apache License 2.0
--------------------------------------------------------------------------------
Copyright (c) 2026 Alibaba Cloud.
Sources:
https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base
MLX:
https://huggingface.co/mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-8bit
https://huggingface.co/mlx-community/Qwen3-TTS-12Hz-1.7B-Base-8bit
Trademark notice: "Qwen" and "Tongyi" are trademarks of Alibaba Innovation
Private Limited. This product uses the Qwen models under Apache License 2.0;
no trademark license is implied beyond reasonable and customary use to
describe the origin of the work (Apache 2.0 § 6).
Qwen3-TTS-Base enables voice cloning. Use of voice cloning is restricted by
the KimiTalk End User License Agreement (Voice Cloning sections), the EU AI
Act Article 50(4) deepfake disclosure obligations, GDPR Article 9 (biometric
data — special category), and personality rights under § 16 ABGB / § 78 UrhG
analog (Austria) and § 823 BGB i.V.m. Art. 1, 2 GG (Germany).
--------------------------------------------------------------------------------
2.4 OmniVoice (k2-fsa / Xiaomi Corp.) Apache License 2.0
--------------------------------------------------------------------------------
Copyright (c) 2026 Xiaomi Corp.
Source: https://huggingface.co/k2-fsa/OmniVoice
OmniVoice code and converted weights are licensed under the Apache License,
Version 2.0. OmniVoice depends on the following sub-components, each with
their own license:
Base foundation model: Qwen3-0.6B
Copyright (c) 2026 Alibaba Cloud Qwen Team
License: Apache License 2.0
Source: https://huggingface.co/Qwen/Qwen3-0.6B
Audio tokenizer: bosonai/higgs-audio-v2-tokenizer
Copyright (c) Boson AI USA, Inc.
License: Boson Higgs Audio 2 Community License Agreement (derived from
Meta Llama 3 Community License Agreement)
Source: https://huggingface.co/bosonai/higgs-audio-v2-tokenizer
Full text: LICENSES/Boson-Higgs-Audio-2-Community-License.txt
Incorporated Meta Llama 3 terms: LICENSES/Meta-Llama-3-Community-License.txt + AUP
--------------------------------------------------------------------------------
2.5 REQUIRED ATTRIBUTION (verbatim English, Boson License § 1.b.i)
--------------------------------------------------------------------------------
Built with Higgs Materials licensed from Boson AI USA, Inc., Copyright Boson AI USA, Inc., All Rights Reserved and Meta Llama 3 licensed under the Meta Llama 3 Community License, Copyright Meta Platforms, Inc., All Right Reserved". based on Meta Llama 3”
--------------------------------------------------------------------------------
2.6 REQUIRED NOTICE (verbatim English, Boson License § 1.b.iii)
--------------------------------------------------------------------------------
“Meta Llama 3 is licensed under the Meta Llama 3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.”
“Boson Higgs Audio 2 is licensed under the Boson Community License, Copyright © Boson AI USA, Inc. All Rights Reserved.”
--------------------------------------------------------------------------------
2.7 ADDITIONAL RESTRICTIONS (Boson License §§ 1.b.iv, 1.b.v, 2)
--------------------------------------------------------------------------------
§ 1.b.iv Meta Llama 3 Acceptable Use Policy compliance is required. Available
at https://www.llama.com/llama3/use-policy — full text reproduced
in LICENSES/Meta-Llama-3-AUP.txt.
§ 1.b.v The Higgs Materials and their outputs must not be used to improve
any other large language model (except Boson Higgs Audio 2 and its
derivatives).
§ 2 Use is permitted only while Licensee's products or services,
including affiliates, do not exceed 100,000 annual active users in
the preceding calendar year. Before that threshold is exceeded, an
expanded license from Boson AI must be granted; otherwise license
rights are suspended unless or until Boson AI expressly grants such
rights. KIMIKON e.U. monitors active-install signals as an early
warning and pauses OmniVoice before the threshold if required for
license compliance.
--------------------------------------------------------------------------------
2.8 Chatterbox Multilingual (Resemble AI) MIT License
--------------------------------------------------------------------------------
Copyright (c) 2025 Resemble AI.
Source: https://huggingface.co/ResembleAI/chatterbox
https://github.com/resemble-ai/chatterbox
MLX: https://huggingface.co/mlx-community/chatterbox-fp16
(converted from ResembleAI/chatterbox via mlx-audio; the conversion
repo's license metadata claims Apache-2.0, which is not supported by
the upstream chain — this product treats the weights as an MIT
derivative and reproduces the upstream MIT notice here)
Full license text: LICENSES/Chatterbox-MIT.txt.
Chatterbox depends on the following runtime sub-component (loaded from the
KIMIKON mirror alongside the main weights):
Speech tokenizer: mlx-community/S3TokenizerV2
Converted from FunAudioLLM/CosyVoice2-0.5B
Copyright (c) Alibaba Group / FunAudioLLM team
License: Apache License 2.0 (see LICENSES/Apache-2.0-FULL.txt)
Source: https://huggingface.co/mlx-community/S3TokenizerV2
This product is not endorsed by or affiliated with Resemble AI. "Chatterbox"
is the upstream project name; no trademark license is implied.
WATERMARK DISCLOSURE: Resemble's reference implementation embeds a PerTh
watermark at generation time. The MLX port used by this Helper does NOT embed
PerTh. AI-audio marking for outputs produced through KimiTalk is performed by
KimiTalk's own mandatory AudioSeal/C2PA provenance pipeline (fail-closed), not
by the model.
Chatterbox supports zero-shot voice cloning from reference audio. Use of voice
cloning is restricted by the KimiTalk End User License Agreement (Voice
Cloning sections), the EU AI Act Article 50(4) deepfake disclosure
obligations, GDPR Article 9 (biometric data — special category), and
personality rights under § 16 ABGB / § 78 UrhG analog (Austria) and § 823 BGB
i.V.m. Art. 1, 2 GG (Germany).
================================================================================
PART 3 — FULL LICENSE TEXTS (LICENSES/ subdirectory)
================================================================================
The adjacent LICENSES/ directory contains the verbatim full text of every
license referenced above:
LICENSES/Apache-2.0-FULL.txt — Apache License 2.0
LICENSES/LGPL-2.1-FULL.txt — LGPL 2.1 (libsndfile)
LICENSES/BSD-2-Clause.txt — BSD 2-Clause (torchaudio)
LICENSES/BSD-3-Clause.txt — BSD 3-Clause (numpy, scipy,
soundfile, PyTorch core)
LICENSES/PSF-Python.txt — Python Software Foundation
LICENSES/Boson-Higgs-Audio-2-Community-License.txt
— Boson custom license,
Meta Llama 3-derived
LICENSES/Meta-Llama-3-Community-License.txt — Meta Llama 3 Community
LICENSES/Meta-Llama-3-AUP.txt — Meta Llama 3 AUP
LICENSES/AudioSeal-MIT.txt — AudioSeal
LICENSES/c2pa-python-MIT.txt — c2pa-python / c2pa-rs binding
LICENSES/Chatterbox-MIT.txt — Chatterbox (Resemble AI)
================================================================================
TRADEMARK NOTICE
================================================================================
KIMITALK, KIMIKON, and the KIMIKON wordmark are trademarks (or unregistered
trademarks pending) of KIMIKON e.U., Vienna, Austria.
All other trademarks are the property of their respective owners. Reasonable
and customary use to describe the origin of bundled software (Apache 2.0 § 6)
is exercised throughout this NOTICE; no trademark license is implied.
================================================================================
END OF NOTICE
================================================================================