[aubio-user] Polyphonics

Martin Pernollet martin.pernollet at gmail.com
Wed Oct 21 15:33:51 CEST 2015


Hi,

Polyphonic transcription of a single instrument is a hard and unsolved
problem!

Although this thesis indicates excellent performances for guitar (98%
recall / 93% precision), they did it in "lab" conditions. The author
honestly admits :

"*These results are also biased by the fact that they were generated using
the same guitar that generated the note templates, and most notes were
played at a constant volume. Playing very quietly or loudly would result in
less accurate note detection. This test reflects the best performance of
our implementation of the NMF algorithm. [...] Unfortunately, the algorithm
performed very poorly on a distorted electric guitar input.*"

Other recent papers [1] focusing on piano transcription and using larger
evaluation datasets announce around 90% precision/recall. Although my
bibliography is not exhaustive, deep learning approaches [2] do not seem
perform better than "classical" machine learning :/

The bad news is that those researchers don't want to share their code... or
sometime lost their code, so one has to implement by her/himself based on
the papers.

Cheers

Martin

[1] http://www.cp.jku.at/research/papers/Boeck_Schedl_ICASSP_2012.pdf
[2]
http://cs.stanford.edu/~jngiam/papers/ismir2011-PolyphonicTranscription.pdf


2015-10-19 23:15 GMT+02:00 <emersont1 at outlook.com>:

>
> Hello
> For polyphonics, I think that this paper, if it has not already been put
> forth, may help (I reuploaded the paper to puush to shorten the URL
> http://puu.sh/kQiCv/c43c603455.pdf).
> I haven't quite had time to read it all, but I think that it would help
> with making aubio polyphonic
> Sent by Outlook <http://taps.io/outlookmobile> for Android
>
> _______________________________________________
> aubio-user mailing list
> aubio-user at aubio.org
> https://lists.aubio.org/listinfo/aubio-user
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.aubio.org/pipermail/aubio-user/attachments/20151021/94998134/attachment.html>


More information about the aubio-user mailing list