With the introduction of whisper.cpp in Debian, it became possible to
enable the Whisper filter in ffmpeg, allowing quick and automatic
transcript of audio using GPU backed audio processing.

Following the recipe available from
<URL: https://medium.com/@vpalmisano/run-whisper-audio-transcriptions-with-one-ffmpeg-command-c6ecda51901f >
I just successfully was able to use the latest ffmpeg from salsa with
the following patch to transcribe the example from that article, ie
using this command:

  ffmpeg -i https://github.com/vpalmisano/webrtcperf/releases/download/videos-1.0/gvr.mp4 -vn -af "whisper=model=$HOME/models-whisper/ggml-model-q5_0.bin\
:language=en\
:queue=3\
:destination=output.srt\
:format=srt" -f null -

The resulting output.srt contained a sensible transcript of the video.

Please consider enabling whisper support in the ffmpeg edition provided
by Debian.

diff --git a/debian/control b/debian/control
index e0fa35ce..2ecf3219 100644
--- a/debian/control
+++ b/debian/control
@@ -190,6 +190,8 @@ Build-Depends:
  libzvbi-dev <!pkg.ffmpeg.stage1>,
 # --enable-opencl
  ocl-icd-opencl-dev | opencl-dev,
+# --enable-whisper
+ libwhisper-dev,
 # used to detect libraries
  pkgconf,
 # HTML documentation
diff --git a/debian/rules b/debian/rules
index 6b3da34f..a2c128f0 100755
--- a/debian/rules
+++ b/debian/rules
@@ -68,6 +68,7 @@ CONFIG := \
        --enable-libxml2 \
        --enable-libxvid \
        --enable-libzimg \
+       --enable-whisper \
        --enable-openal \
        --enable-opencl \
        --enable-opengl

Control: tags -1 wontfix

Not at the moment. whisper.cpp and its (build-)dependencies are not in
testing and looking at their current state they are not ready to be
promoted to key packages.

Cheers

Sebastian Ramacher wrote:

whisper.cpp migrated to testing 2 weeks ago.

Please reconsider enabling support for it here.

Kind regards,

 - Jonas

It does not look like whipser.cpp and its (transitive) dependencies are
ready to become key packages:

 Marked for autoremoval on 05 September due to aardvark-dns, golang-github-bmizerany-assert, golang-github-masterminds-sprig, httpcore, jupyter-notebook, node-playwright, node-yarnpkg, towncrier, wasmedge: #1133663, #1135849, #1138720, #1141793, #1143319, #1143321, #1143428, #1143887, #1143916, #1143952

Cheers

Hi Sebastian,

Quoting Sebastian Ramacher (2026-08-16 13:40:26)
month" is sensible. I suspect ffmpeg would not fit that bar - but as I
understand it we may never know that, because autoremovals are computed
only for non-key packages.

Put another way: What do ffmpeg maintainers consider the threshold for
acceptance, if neither "has stayed in testing for 2 week" nor "has
stayed testing for 4 months"?

 - Jonas

Hi Jonas

Well, we do know and I have removed problematic dependencies from ffmpeg
to get them out of the key packages set in the past.

As we (as in the release team) are trying to reduce the size of the
key package set, I am not willingly adding new packages to that set that
have RC bugs open and unanswered for 4 months.

Cheers

Quoting Sebastian Ramacher (2026-08-16 14:56:31)

Oh! I was unaware that this issue involves RC bugs ignored for 4 months
and agree that is a sensible bar to set. Which bugs is it?

Kind regards,

 - Jonas

ICYMI, since the original request last February, ffmpeg support has
been enabled in the whisper.cpp package itself.

This was introduced mid-June in whisper.cpp 1.8.6+dfsg-2.

We hadn't enabled it sooner because the licensing of the related code
in whisper.cpp was not clear. We raised it to upstream, who kindly
rewrote this part [1].

So, it means that the requested feature could already be used the other
way around: using whisper.cpp to transcribe from ffmpeg-supported
formats.

Given the release constraints raised in this bug, it seems to me that
we should rather recommend this usage (whisper.cpp using ffmpeg),
rather than having ffmpeg depending on whisper.cpp (which, while more
stable than llama.cpp, is still a relatively recent library).

Anyhow, the potential cyclic dependency should may be also be
considered, as the whisper.cpp package now depends on ffmpeg.
that
months

Other people in the AI team know these issues much better, but my
understanding is that many such bugs are usually coming from the main
whisper.cpp dependency, ggml (also used by llama.cpp), and especially
some of its backends (ROCm).

If there is an issue specifically with whisper.cpp that we may have
missed, please let me know or raise it in the bug tracking system.

Cheers,

Mathieu

[1] https://github.com/ggml-org/whisper.cpp/issues/3838