extracting audio features

[Avishek Saha]

Hi All,

I am working on a recommender system and was planning to extract an MSD dataset of the following form:

USER   LABEL (listened or not listened)  MUSIC_FEATURES 

I can extract the USER-TRACK information from the Taste Profile dataset which would give me the labels.

My question is about the features:

 -- I know I can extract lyrics from MusicMatch

 -- The audiowave features like beat, loudness, etc are available in the MSD database. But there are too many of them and since I don't have much background on music research, I was wondering which ones to use. For example, I would be happy with the feature set in the YearPrediction data -- but the trackid is missing here and so I can match these features with the other datasets.

Has someone faced a similar situation before?

Does the MSD team already have a dataset in the above format? If yes, then it would be great if it could be made available on the MSD and/or UCI repo. This could be very useful for machine learning applications like multitask learning where the need of the hour is a large scale dataset in the above format and MSD seems to perfectly fit the bill. If the data is not already available then kindly let me know how to match the data in the YearPrediction to track ids of maybe how to extract an audio feature subset from the database.

Thanks,

Avishek

[Thierry Bertin-Mahieux]

Hi,
the problem is that there is zero consensus on how to best summarize The Echo Nest features. This is an active research question (even saw some abstracts on the subject at the last ISMIR bidding).
For the year prediction task, we did the simplest, maybe stupid transformation we could think of, using mean and covariance coefficient (and ignoring the different lengths of the songs, the varying time-scale of the frames, the varying loudness, etc).
We don't maintain any "projected collections" of tid -> features, even if we tried a few in our research.
I can send you the exact code we used to create the features on the UCI repository from the original HDF5 files, but you could re-code it in 5 min and it does required having the full MSD at hand.
Sorry we can't help you more, I would love to see the community define a common set of projected features (low-dimensional) for quick experimentation, but it doesn't exist yet.
Thierry

--
You received this message because you are subscribed to the Google
Groups "Million Song Dataset" group.
To post to this group, send email to millionso...@googlegroups.com
To unsubscribe from this group, send email to
millionsongdata...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/millionsongdataset?hl=en?hl=en
Million Song Dataset main webpage:
http://labrosa.ee.columbia.edu/millionsong/

--
Thierry Bertin-Mahieux
Columbia University, NY
Ph.D. Candidate, Electrical Engineering
http://www.columbia.edu/~tb2332/