[prev in list] [next in list] [prev in thread] [next in thread] 

List:       wekalist
Subject:    Re: [Wekalist] Help with song classification project
From:       Sione <sionep () xtra ! co ! nz>
Date:       2006-10-28 22:50:24
Message-ID: 4543DEB0.90201 () xtra ! co ! nz
[Download RAW message or body]

wekalist-request@list.scms.waikato.ac.nz wrote:

> The main idea of my project is to classify objects (songs) from a 
> collection
> through a few attributes that describes them.
> 
> Actually, I´m building an experiment to prove (or test) if my idea is
> possible and I´m familiar with the experimental tutorial. However, I facing
> some problems to get clear results on my task.
> 
> Here is the deal. I got a arff file with a hundred songs, each one is
> described by nine attributes (like tempo, rhythm, etc.). I wish I could
> retrieve a list of relevance by similarity based on given attributes (an
> easily task for KNN algorithm).
> 
> Now, follow a list of problems that I found using the Weka experimenter:
> 
> 1) In the arff file, I can´t use unnormalized fields for the 1BK algorithm
> and there are two fields for this situation: artist name e song name. I
> considered them irrelevant, because it identifies the songs that I´m
> classifying. Until now, my solution was to describe all songs e artist as
> class attributes.
> 
> 2) I can´t have a list of songs, it just tells me if the select song was
> correctly classified or not. It forces me to create a goal field, that is
> not my main intention.
> 
If you want to do  song classification more accurately using signal processing \
techniques in combination with machine learning then I think that you might be \
interested to look at some papers found here: 

"Music Technology Group (MTG)"
http://www.iua.upf.edu/mtg/publicacions.php

There are commercial music recognition products originated from MTG which are license \
to BMat Ltd (see link below).

"BMat"
http://www.bmat.com/technologies.html 

Here is another music recognition commercial product.

"DoubleV3"
http://www.doublev3.com/index_en.html

I was going to develop something similar a while ago , by using  DSP (digital signal \
processing) algorithms in combination with machine learning, but at the end, the \
company who wanted this "music recognition recommender systems" developed inhouse \
opted for a commercial tool so that they can hit the ground running.  

DSP algorithms will decompose songs into their spectral features mostly by  STFT \
(short-time-fourier-transform). This data can be used as input to the machine \
learning algorithms. Both commercial products stated above use DWT (discrete wavelet \
transform) for song feature extractions. One of them do use ICA for source \
separation, that is a song can be decomposed entirely into its different components \
signals. Eg, the lead singer's voice could be completely extracted from the song or \
the bass guitar , etc,... There are some open source Java wavelet on the internet , \
just do a search.

Cheers,
Sione.




_______________________________________________
Wekalist mailing list
Wekalist@list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist


[prev in list] [next in list] [prev in thread] [next in thread]