 |
Musical Artist Recognition
I used the music similarity measure described in Aucouturier and Pachet's Music Similarity Measures: What's the Use? to build an artist classifier described below.
Data
First, I collected a small database of music containing 10 artists and 514 songs:
| Artist | Number of Songs |
| Beastie Boys | 90 |
| Cranberries | 52 |
| Dave Matthews Band | 56 |
| Led Zeppelin | 54 |
| Massive Attack | 28 |
| Material Issue | 38 |
| Miles Davis | 10 |
| Nine Inch Nails | 78 |
| Pearl Jam | 80 |
| Tool | 28 |
Features
I converted each song to a single channel (mono) sampled at 22050 Hz. Then, I segmented the audio into non-overlapping frames of 1024 samples. I extracted the first 8 mel-frequency cepstral coeffiects (MFCCs) from each frame resulting in approximately 4000 feature vectors per song (32000 coefficients).
Model
I modeled each song as a mixture of three Gaussian distributions. That is, I interpret the feature vectors as Gaussian random variables drawn from a probability density function (pdf). I presuppose that the pdf is reasonably represented as the sum of three Gaussian distributions and estimate the means and variances of them using the data from each song. An example of a Gaussian mixture model (GMM) with three Gaussians is shown below:
Test Set
I used half of the songs from each artist to train a GMM that represents each artist. In order to classify each of the remaining songs, I computed the likelihood that the song's feature vectors were generated by each of the artist models. The confusion matrix below shows the number of songs from the artist in that row that were classified as the artist in the column. For example, 29 Beastie Boys songs were classified correctly but 5 were classified as Massive Attack and 4 as Nine Inch Nails.
| Artist | BB | CR | DM | LZ | MA | MD | MI | NI | PJ | TO | Sum |
| Beastie Boys | 29 | 0 | 0 | 1 | 5 | 2 | 2 | 4 | 2 | 0 | 45 |
| Cranberries | 1 | 17 | 0 | 0 | 2 | 1 | 3 | 0 | 2 | 0 | 26 |
| Dave Matthews Band | 0 | 0 | 23 | 2 | 1 | 0 | 0 | 1 | 1 | 0 | 28 |
| Led Zeppelin | 1 | 2 | 3 | 12 | 0 | 0 | 5 | 0 | 3 | 1 | 27 |
| Massive Attack | 1 | 0 | 0 | 0 | 12 | 0 | 0 | 1 | 0 | 0 | 14 |
| Miles Davis | 1 | 0 | 0 | 0 | 1 | 3 | 0 | 0 | 0 | 0 | 5 |
| Material Issue | 0 | 1 | 0 | 1 | 0 | 0 | 16 | 0 | 1 | 0 | 19 |
| Nine Inch Nails | 1 | 1 | 1 | 1 | 3 | 1 | 0 | 28 | 0 | 3 | 39 |
| Pearl Jam | 1 | 3 | 1 | 6 | 5 | 0 | 11 | 0 | 13 | 0 | 40 |
| Tool | 0 | 0 | 1 | 0 | 1 | 0 | 0 | 2 | 0 | 10 | 14 |
| Total | 35 | 24 | 29 | 23 | 30 | 7 | 37 | 36 | 22 | 14 | 257 |
Last modified 28 October 2009 at 10:06 am by Mitch Parry |
