Unicorns Can TalkThird-party notices
Unicorns Can Talk is released under the BSD 3-Clause License (see LICENSE). That licence covers this program's own code, its unicorn artwork, its skins and its documentation. The program also uses, and its Windows build includes, other people's software, and (once you use them) speech and language models that it downloads. Each of those keeps its own licence, listed here. The full licence text of each bundled Python package is in the licenses folder next to the program.
This list was written from the packages' own metadata and the build's own files on 2026-11-06. It is a good-faith summary, not legal advice. If you redistribute the program, check it against what you actually ship.
1. Libraries included in the program
| Component | Licence | What it is used for |
|---|---|---|
| Python (PSF) | Python Software Foundation License | the runtime inside the program |
| PySide6 / Qt for Python, Shiboken6 (The Qt Company) | LGPL-3.0 (or GPL-2.0 / GPL-3.0, or commercial) | all windows and the tray; see section 2 |
| numpy | BSD-3-Clause (and 0BSD, MIT, Zlib, CC0 parts) | audio and number arrays |
| sherpa-onnx, sherpa-onnx-core (k2-fsa) | Apache-2.0 | the Parakeet speech model and voice detection |
| onnxruntime (Microsoft) | MIT | runs ONNX models |
| ctranslate2 (OpenNMT) | MIT | runs the Whisper speech models |
| faster-whisper (Systran) | MIT | Whisper speech models |
| llama-cpp-python (and llama.cpp) | MIT | runs the proofreading language model |
PyAV (av) | BSD-3-Clause, with FFmpeg inside; see section 3 | reads audio files for transcription |
| soundfile (and libsndfile) | BSD-3-Clause (libsndfile: LGPL-2.1) | reads and writes audio files |
| sounddevice (and PortAudio) | MIT (PortAudio: MIT-style) | the microphone |
| PyYAML | MIT | settings files |
| psutil | BSD-3-Clause | process and hardware information |
| certifi | MPL-2.0 | certificates for downloads |
| huggingface_hub, hf-xet, tokenizers | Apache-2.0 | downloading and reading models |
| pywin32 | PSF | talking to Microsoft Word |
| breame | Apache-2.0 | British and American spelling |
| Jinja2, MarkupSafe, click, httpx/httpcore, tqdm, filelock, fsspec, packaging, typing_extensions, cffi, protobuf, diskcache, flatbuffers | BSD-3-Clause, MIT, MPL-2.0, Apache-2.0, PSF (as listed in licenses) | supporting libraries of the above |
| PyInstaller (bootloader) | GPL-2.0-or-later with the special exception that allows the programs it builds to use any licence | builds UnicornsCanTalk.exe |
2. Qt (PySide6) and the LGPL
Qt for Python is used under the GNU Lesser General Public License, version 3. The Qt libraries are included unmodified and dynamically linked: they are separate files in the program's _internal\PySide6 folder, and you may replace them with other builds of the same Qt version. Qt's source code is available from https://www.qt.io/download-open-source and https://code.qt.io, and the licence text from https://www.gnu.org/licenses/lgpl-3.0.html. This program is not endorsed by The Qt Company.
3. FFmpeg and the codecs inside PyAV
PyAV's Windows build carries its own copy of FFmpeg (https://ffmpeg.org) in the av.libs folder. That FFmpeg identifies itself as LGPL version 3 or later and was not built with "non-free" components. The same folder also contains these codec and support libraries:
| Library | Licence |
|---|---|
| libx264 | GPL-2.0-or-later |
| libx265 | GPL-2.0-or-later |
| libdav1d, libopus, libvpx, libwebp, libvmaf | BSD-2-Clause / BSD-3-Clause |
| SVT-AV1 | BSD-3-Clause-Clear (with AOM patent licence) |
| libvpl | MIT |
| libiconv, LAME (MP3) | LGPL-2.1 / LGPL-2.0 |
| MinGW runtime (libgcc, libstdc++, winpthreads), zlib | GPL-3.0 with runtime exception, MIT, zlib |
This program only reads (decodes) audio files for the "Transcribe a recording" feature. It never encodes video and never uses x264 or x265. Even so, those two GPL libraries are present in the bundle. If you redistribute the Windows program to others, you are responsible for complying with the licences above, including making the corresponding source available on request: FFmpeg https://ffmpeg.org/download.html, x264 https://www.videolan.org/developers/x264.html, x265 https://x265.org, PyAV https://github.com/PyAV-Org/PyAV. For a wide public release, take advice, or replace PyAV's audio reading with an LGPL-only route (this is on the roadmap).
4. Speech and language models (downloaded when you first use them)
These are not part of the installer. The program downloads the one you choose (or that suits your computer) from their publishers' repositories on Hugging Face, and each keeps its own licence.
| Model | Publisher | Licence | Notes |
|---|---|---|---|
| Parakeet-TDT 0.6B v2 (ONNX int8 conversion by csukuangfj / k2-fsa) | NVIDIA | CC BY 4.0 | the default speech model. Credit: "Parakeet-TDT 0.6B v2 by NVIDIA, licensed CC BY 4.0, https://creativecommons.org/licenses/by/4.0/; ONNX conversion by k2-fsa (sherpa-onnx)" |
| Whisper base.en, small.en, large-v3-turbo; distil-large-v3 (faster-whisper conversions) | OpenAI; Hugging Face (distil); Systran / mobiuslabs (conversion) | MIT | optional speech models |
| Silero VAD | Silero Team | MIT | finds the pauses between phrases |
| Qwen2.5 1.5B Instruct (GGUF) | Alibaba Cloud (Qwen) | Apache-2.0 | proofreading, small |
| Qwen2.5 3B Instruct (GGUF) | Alibaba Cloud (Qwen) | Qwen Research License: non-commercial / research use only | proofreading, the default on most PCs; https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/blob/main/LICENSE. For commercial use choose the 1.5B or 7B model in Settings > Proofreading. |
| Qwen2.5 7B Instruct (GGUF) | Alibaba Cloud (Qwen) | Apache-2.0 | proofreading, most careful, slow without a GPU |
The BSD 3-Clause licence of this program does not change a model's licence: the licence of the model you download applies to your use of that model. Settings > Proofreading and the download prompt say which licence each proofreading model has.
5. Artwork and fonts
The unicorn artwork (face, icon, installer pictures, the five skins) was drawn for this project and is covered by the BSD 3-Clause licence. "Unicorns Can Talk" and the unicorn face are not a trademark grant: clause 3 of the licence still means you may not use the name to endorse a product derived from this one without permission. No fonts are bundled; the windows use fonts already on your computer (Segoe UI, Consolas).
6. Where to find the full texts
the licence files that package ships).
- Bundled Python packages: the
licensesfolder next to the program (one folder per package, with - This program:
LICENSE(BSD 3-Clause), also at https://unicorns-can.talk/license. - These notices: also at https://unicorns-can.talk/notices.