Inside Collection (Course): ELEC 301 Projects Fall 2006
Summary: General introduction of the LPC approach of our genre detection method.
The LPC approach we used was based on the Linear Predictive Coefficients, which is a function in MATLAB. After researching predictive models, we determined that the LPC, which is normally applicable to speech, could be used for sound prediction as well. It works on the premise that given a vector, the LPC function will determine the coefficients of an Nth order forward linear predictor.
A = LPC(X,N)
A = [1 A (2) A (3) ... A (N+1) ]
We used the dot product to find the error between the predicted nth term and the actual nth term of the input music. For our purposes, we chose to use the 25th order after determining that it yielded the highest accuracy experimentally.
Error = A * X
= X (n) – A (2)*X (n-1) – A (3)*X (n-2) – ... – A (N+1)*X (n-N)
The predictor was used to determine the error between the input song and the genre (our genre database). The idea is that the genre that gives the least error would be the genre of the input song.
RockLPCTotal ErrorClassicalLPCTotal ErrorRapLPCTotal ErrorInput Song? Genre ?
This figure represents the testing of the song with each genre database. Testing consisted of the following procedure:
We first loaded the input song. Then applied it to the database for each genre (one at a time) to determine which genre it is most similar to. For the first genre, find the LPC coefficients for it. Then section the input song vector into two sections of length 25 each, as shown in figure 2. Apply each section to the LPC coefficients to find the error for each section. Then sum up the errors in each section to find the total error for that genre. Finally, do it for all the genres. By the end of this procedure, we would have the total errors for each genre. And the minimum error should be the genre that the input song belonged to.
Sections of song vector (fig. 2)
We found several specific errors in using the LPC method. First and foremost, it took too much time to calculate the errors (about 53 seconds per input song). Thus, we could not be able to put enough sampled songs into our database, which lead to get more accurate results. However, if we used another method of forming our database, specifically by using a matrix, the computations would have been much faster, and we could have used more songs in our database. The problem was that in our program, we recalculated the LPC coefficients each time we ran checked an input song. Had we saved the LPC coefficient values in a matrix, the computations would have been much faster.
The LPC could not accurately detect neither jazz nor pop under the algorithm we had developed. In fact, the LPC was most accurate when deciding between two random types of songs, but not five. It was particularly accurate with rap and classical.
More about this module: Metadata | Downloads | Version History
This work is licensed by Yiming Wang under a Creative Commons Attribution License (CC-BY 2.0), and is an Open Educational Resource.
Last edited by Rice University ELEC 301 on Sep 25, 2007 2:08 pm -0500.
