Docsity
Docsity

Prepare for your exams
Prepare for your exams

Study with the several resources on Docsity


Earn points to download
Earn points to download

Earn points by helping other students or get them with a premium plan


Guidelines and tips
Guidelines and tips

Assignment 4: Mel-Frequency Cepstral Coefficients & Nearest Neighbor in Music Retrieval, Exercises of Music Information Retrieval

An assignment for the music information retrieval course, where students are required to write a matlab function to extract mel-frequency cepstral coefficients (mfcc) from a given 16-bit wav file sampled at 44.1khz, create non-overlapping 1-second-long subsequences, compute average mfcc vectors, and use a nearest neighbor classifier to label each subsequence based on a predefined set of 4 music instrument classes. Students are also asked to create training and testing sets, report classification accuracy, and discuss the impact of processing parameters and number of neighbors and classes on the results.

Typology: Exercises

2011/2012

Uploaded on 01/25/2012

paulina
paulina 🇺🇸

4.4

(13)

241 documents

1 / 1

Toggle sidebar

This page cannot be seen from the preview

Don't miss anything!

bg1
MPATE-GE 2623 Music Information Retrieval - Assignment # 4
Instructor: Juan Pablo Bello
Please check the course’s website for detailed submission instructions.
Write a Matlab function that takes the name of any 16-bit wav file (if stereo, please simply
disregard the 2nd channel) sampled at fs= 44.1kHz , and performs the following operations:
computes the sequence of Mel-frequency cepstral coefficients (MFCC).
breaks the feature sequence into 1 second-long, non-overlapping subsequences, and
computes the average MFCC vector for each subsequence.
uses a nearest neighbor classifier (using the Euclidean distance) to label each subse-
quence as belonging to one of a predefined set of 4 music instrument classes.
plots the output of each operation as subplots in a single figure.
Please note that you will need to create both “training” and “testing” sets of manually-
labeled audio files evenly distributed across the 4 classes (e.g. 50 seconds each of violin,
flute, piano and trombone). It is recommended that you select instrument classes that can
be easily differentiated from each other. Train and test files should be recordings of a single
instrument (or a concatenation of excerpts featuring only one instrument at a time). Note
that recordings of the same instrument should be different for each set.
Please report the percentage of 1-second windows that are correctly classified, per in-
strument class and overall. Include graphs in your report as necessary. Comment on the
resulting plots, their differences and their relationship to the characteristics of the sounds,
and your choice of processing parameters.
Extra point: Experiment with increasing the number of nearest neighbors and of instrumental
classes. Discuss how these modifications impact classification.
1

Partial preview of the text

Download Assignment 4: Mel-Frequency Cepstral Coefficients & Nearest Neighbor in Music Retrieval and more Exercises Music Information Retrieval in PDF only on Docsity!

MPATE-GE 2623 Music Information Retrieval - Assignment # 4 Instructor: Juan Pablo Bello

Please check the course’s website for detailed submission instructions.

Write a Matlab function that takes the name of any 16-bit wav file (if stereo, please simply disregard the 2nd channel) sampled at fs = 44. 1 kHz, and performs the following operations:

  • computes the sequence of Mel-frequency cepstral coefficients (MFCC).
  • breaks the feature sequence into 1 second-long, non-overlapping subsequences, and computes the average MFCC vector for each subsequence.
  • uses a nearest neighbor classifier (using the Euclidean distance) to label each subse- quence as belonging to one of a predefined set of 4 music instrument classes.
  • plots the output of each operation as subplots in a single figure.

Please note that you will need to create both “training” and “testing” sets of manually- labeled audio files evenly distributed across the 4 classes (e.g. 50 seconds each of violin, flute, piano and trombone). It is recommended that you select instrument classes that can be easily differentiated from each other. Train and test files should be recordings of a single instrument (or a concatenation of excerpts featuring only one instrument at a time). Note that recordings of the same instrument should be different for each set. Please report the percentage of 1-second windows that are correctly classified, per in- strument class and overall. Include graphs in your report as necessary. Comment on the resulting plots, their differences and their relationship to the characteristics of the sounds, and your choice of processing parameters.

Extra point: Experiment with increasing the number of nearest neighbors and of instrumental classes. Discuss how these modifications impact classification.