[FFmpeg-devel] [OPW] OPW Project Proposal

[Pallavi Kumari]

Thanks for the inspiring reply Nicholas :) I am glad that you find the idea
very interesting.

>> I think this could be a very interesting project, but maybe also very
hard.
>> I myself do not have the skills or knowledge to help in any way in that
area.

I have been reading few papers related to such field and have some basic
idea as my research interest is in the area of information retrieval, ML
and Deep learning. I think this could be doable. Also, looking forward for
mentor(s) for this project.

>> OTOH, there is one use that I have in mind that may not be covered by
existing algorithms: fingerprinting a short
>> segment of audio to search for it efficiently and accurately in a larger
footage without access to the original
>> segment, and even if it was subjected to a different lossy codec or even
small changes in pitch or speed. To be
>> honest, I do not know if it is even possible, but it would be very
useful.

This was something in my mind too. Music application like Shazam does a
pretty good job for this purpose.  Given a piece of music from an audio (or
humming manual) it finds exact song and the similar songs.

Algorithms for the above scenario can be done by doing spectrum analysis
(fingerprinting) technique. Let me make a draft explaining algorithmic
steps in details.

Looking forward to more interesting views/opinions.


Regards,
Pallavi,
IRC - atana

On Mon, Sep 26, 2016 at 10:33 PM, Nicolas George <george at nsup.org> wrote:

> Le quartidi 4 vendémiaire, an CCXXV, Pallavi Kumari a écrit :
> > I want to propose the idea of implementing filters for ffmpeg that would
> > give different audio fingerprints for an audio which could be reused by
> > other people for variety of applications. Goal of this system is given a
> > song as a input, it would spits out similar sounding songs. Some of its
> > applications are:
> >
> > * detecting duplicate songs
> > * retrieving similar songs
> > * can be used for recommending songs
> > * clustering songs based on content/tags
>
> > Kindly, let me know if the proposed idea sounds interesting and relevant
> to
> > the community. I would be happy to share a rough draft containing
> details.
>
> I think this could be a very interesting project, but maybe also very hard.
> I myself do not have the skills or knowledge to help in any way in that
> area.
>
> I think a few fingerprinting algorithms of the kind exist. Implementing
> them
> with ffmpeg's DSP infrastructure would probably be easier than designing
> one
> from scratch.
>
> OTOH, there is one use that I have in mind that may not be covered by
> existing algorithms: fingerprinting a short segment of audio to search for
> it efficiently and accurately in a larger footage without access to the
> original segment, and even if it was subjected to a different lossy codec
> or
> even small changes in pitch or speed. To be honest, I do not know if it is
> even possible, but it would be very useful.
>
> Regards,
>
> --
>   Nicolas George
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
>