- ubuntu 16.04 LTS
- python3.5.2 (using Pycharm 2018.1)
- extra modules: numpy, scipy, matplotlib, prettytable, librosa, keras
Before using the codes, please download GTZAN dataset and it's key notation; then go to utils.py, change audio_dir and label_dir to the location where you download the data.
audio_dir = '/your/dataset/path'
label_dir = '/your/dataset/anotation/path'
In this task, we use binary major/minor key template to compare the similarity with summed chroma feature. Only test on five genre.
bin_major_template = np.array([[1, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1]]) / np.sqrt(7.0)
bin_minor_template = np.array([[1, 0, 1, 1, 0, 1, 0, 1, 1, 0, 1, 0]]) / np.sqrt(7.0)
The program will output prediction of each song, genre raw accuracy and MIREX accuracy. The MIREX accuracy procedure is described here.
+--------+------+---------------+----------------------+----------------------+----------+----------------+
| Genre | Same | Perfect Fifth | Relative Minor/Major | Parallel Minor/Major | Accuracy | MIREX Accuracy |
+--------+------+---------------+----------------------+----------------------+----------+----------------+
| pop | 37 | 21 | 7 | 9 | 0.393617 | 0.546809 |
| blues | 7 | 10 | 3 | 46 | 0.071429 | 0.225510 |
| metal | 24 | 17 | 3 | 21 | 0.258065 | 0.404301 |
| hiphop | 10 | 4 | 5 | 11 | 0.123457 | 0.193827 |
| rock | 39 | 22 | 8 | 8 | 0.397959 | 0.551020 |
| All | 117 | 74 | 26 | 95 | 0.252155 | 0.389655 |
+--------+------+---------------+----------------------+----------------------+----------+----------------+
You can try to change the value g in Q1_3.py, which is the factor of logarithmic compression, to see how it affects the result.
The task is quite the same as the previous one, but we change binary template to Krumhansl-Schmuckler profile, which highly enhance the detection result.
ks_major_template = np.array([[6.35, 2.23, 3.48, 2.33, 4.38, 4.09, 2.52, 5.19, 2.39, 3.66, 2.29, 2.88]])
ks_minor_template = np.array([[6.33, 2.68, 3.52, 5.38, 2.6, 3.53, 2.54, 4.75, 3.98, 2.69, 3.34, 3.17]])
+--------+------+---------------+----------------------+----------------------+----------+----------------+
| Genre | Same | Perfect Fifth | Relative Minor/Major | Parallel Minor/Major | Accuracy | MIREX Accuracy |
+--------+------+---------------+----------------------+----------------------+----------+----------------+
| pop | 59 | 12 | 4 | 6 | 0.627660 | 0.717021 |
| blues | 15 | 1 | 1 | 50 | 0.153061 | 0.263265 |
| metal | 38 | 9 | 2 | 20 | 0.408602 | 0.506452 |
| hiphop | 20 | 10 | 4 | 9 | 0.246914 | 0.345679 |
| rock | 48 | 23 | 6 | 8 | 0.489796 | 0.641837 |
| All | 180 | 55 | 17 | 93 | 0.387931 | 0.498276 |
+--------+------+---------------+----------------------+----------------------+----------+----------------+
This task actually ask students to design a new algorithm to perform key finding, and I found that blues has the lowest score among the above method. So I decided to add two extra blues major/minor template to the KS-profile method, and the result outperform the previous method overall.
blues_major_template = np.array([[1, 0, 1, 1, 1, 0, 0, 1, 0, 1, 0, 0]]) / np.sqrt(6)
blues_minor_template = np.array([[1, 0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 0]]) / np.sqrt(6)
+--------+------+---------------+----------------------+----------------------+----------+----------------+
| Genre | Same | Perfect Fifth | Relative Minor/Major | Parallel Minor/Major | Accuracy | MIREX Accuracy |
+--------+------+---------------+----------------------+----------------------+----------+----------------+
| pop | 57 | 11 | 5 | 5 | 0.606383 | 0.691489 |
| blues | 26 | 5 | 1 | 43 | 0.265306 | 0.381633 |
| metal | 39 | 10 | 4 | 18 | 0.419355 | 0.524731 |
| hiphop | 21 | 10 | 5 | 3 | 0.259259 | 0.346914 |
| rock | 48 | 20 | 10 | 11 | 0.489796 | 0.644898 |
| All | 191 | 56 | 25 | 80 | 0.411638 | 0.522629 |
+--------+------+---------------+----------------------+----------------------+----------+----------------+
Different from the above task, we need to find the key at each time steps in the songs. I use KS-profile matching Q4 and lots of average filtering(window size ~= 64s) on the chroma feature, eventually reach almost 60% of MIREX accuracy.
The data used here is Beethoven’s Piano Sonata Functional Harmony made by the course teacher, unfortunately I can't give it to others.
song accuracy mirex accuracy
1.wav 0.517500 0.642500
3.wav 0.358993 0.578058
5.wav 0.427677 0.490403
6.wav 0.298267 0.494183
8.wav 0.608796 0.661632
11.wav 0.657303 0.723876
12.wav 0.301370 0.548493
13.wav 0.364103 0.482051
14.wav 0.666667 0.731522
16.wav 0.269495 0.428211
18.wav 0.757812 0.798437
19.wav 0.665517 0.739655
20.wav 0.216867 0.328112
21.wav 0.481338 0.555470
22.wav 0.623423 0.697838
23.wav 0.358298 0.432295
24.wav 0.571078 0.629534
25.wav 0.431900 0.612455
26.wav 0.504167 0.683417
27.wav 0.676190 0.740544
28.wav 0.558824 0.653595
31.wav 0.658046 0.696264
32.wav 0.626794 0.719019
overall 0.488095 0.594645
I also tried using CNN to solve the task, you can check out here.