License of whisper.cpp with FFMPEG support?

[petterreinholdtsen]

A question have come up in the proposal to add FFMPEG support for the version of whisper.cpp in Debian, see <URL: https://salsa.debian.org/deeplearning-team/whisper.cpp/-/merge_requests/10 > for the background. The question is how the license of
examples/ffmpeg-transcode.cpp will affect the license of programs linking to whisper.cpp, and if the license of this file is correct.

In PR #2133 the file examples/ffmpeg-transcode.cpp was added to the code base. This file is marked with SPDX-License-Identifier: GPL-2.0, while README.md claim the whisper.cpp license is MIT. Given the terms of these two styles of licensing, I believe the general concensus is that the combined work will end up with GPL-2.0 license. This might be problematic for
programs with GPL-incompatible licenses that are compatible with MIT.

If the file really is GPL-2.0 licensed, would it be possible to relicense it with MIT license, or at least with GPL 2.0 or later at the users discretion?

According to the file header, the copyright holders are Andrew Clayton andrew@digital-domain.net (2019) and William Tambellini william.tambellini@gmail.com (2024).

A complicating factor is that ffmpeg itself is licensed according to GNU Lesser General Public License version 2.1 or later (LGPL v2.1+).

[WyattBlue]

It would make more sense to compile ffmpeg with whisper.cpp support instead of the other way around.

[ckastner]

@mbaudier pointed out an important point: it looks as if this affects only the utility; the library should remain pure MIT.

The distinction is important for third-party utilities linking to libwhisper (which I assume is a common use case) and where the GPL's position is that linking to GPLed code alone already triggers the GPL's viral nature.

Nevertheless, if the two authors listed by @petterreinholdtsen were to agree to relicense under the GPL MIT, that would certainly be the easiest outcome.

[WyattBlue]

Nevertheless, if the two authors listed by @petterreinholdtsen were to agree to relicense under the GPL, that would certainly be the easiest outcome.

Pure nonsense unless you mean "GPLv2 or later". I would prefer "LGPLv2.1 or later". Regardless, why is the utility binary compiling and linking to an example file anyway?

[mbaudier]

Nevertheless, if the two authors listed by @petterreinholdtsen were to agree to relicense under the GPL, that would certainly be the easiest outcome.

You meant: "to relicense under the MIT license", right?

[mbaudier]

Regardless, why is the utility binary compiling and linking to an example file anyway?

@WyattBlue In whisper.cpp, the sources for the utilities are under ./examples (e.g. ./examples/cli for whisper-cli)

[WyattBlue]

@mbaudier Okay, this project needs to document that the utility is under 'GPLv2 only', even if the library is MIT

[ckastner]

Nevertheless, if the two authors listed by @petterreinholdtsen were to agree to relicense under the GPL, that would certainly be the easiest outcome.

You meant: "to relicense under the MIT license", right?

Ah yes, thanks. I fixed my comment.

[mbaudier]

@mbaudier Okay, this project needs to document that the utility is under 'GPLv2 only', even if the library is MIT

Only if the utilities are built with ffmpeg support. Which is why we (Debian package co-maintainers) don't enable ffmpeg support for the time being, until everything is clarified.

I personally use the library directly from Java, and the sound handling is done with standard javax.sound.* utilities which support only WAV anyway. This would be handy to have the tools supporting other formats while experimenting though.

[ggerganov]

With v1.8.6 onwards there is no longer GPL code in the examples - see #3846 for more information.