yoyololicon / mir_hw2

Homework 2 for Music Information Retrieval

Environment

• ubuntu 16.04 LTS
• python3.5.2 (using Pycharm 2018.1.4)
• extra modules: numpy, scipy, matplotlib, prettytable, librosa

Dataset

All the experiments were done on Ballroom dataset, consist of 30s long ballroom dancing music pieces.

How to use ?

1. Download the raw audio files.

2. Download the tempo annotations of each pieces.

3. Donwload the beat annotations.

4. Modified the directory variables in utils.py to where you unzip the data.

   data_dir = '/where/you/put/audio/files'
   bpm_label_dir = '/where/you/put/bpm/annotations'
   beat_label_dir = /where/you/put/beat/annotations'
   

Usage of each file

Q1~3: Tempo estimation using Fourier tempogram

In this task, we use Fourier tempogram, which is the short-time Fourier transform of the spetral flux novelty curve of the music, to perform tempo estimation.

$ python3 Q1-3.py

The program will output its P-score and ALOTC score (at least one tempo correct) on eight different genres.

+---------------+---------+--------+--------------+--------------+--------------+
|     Genre     | P-score | ALOTC  | 1/2T P-score | 1/3T P-score | 1/4T P-score |
+---------------+---------+--------+--------------+--------------+--------------+
|   ChaChaCha   |  0.3285 | 0.7387 |    0.4968    |    0.0085    |    0.1571    |
|      Jive     |  0.4543 | 0.8333 |    0.1583    |    0.0437    |    0.0131    |
|   Quickstep   |  0.4785 | 0.9146 |    0.0085    |    0.0000    |    0.0000    |
|     Rumba     |  0.2185 | 0.4898 |    0.5141    |    0.0083    |    0.2032    |
|     Samba     |  0.1332 | 0.2907 |    0.4468    |    0.0000    |    0.3912    |
|     Tango     |  0.4958 | 0.9535 |    0.3841    |    0.0000    |    0.0000    |
| VienneseWaltz |  0.5323 | 0.9692 |    0.0854    |    0.0000    |    0.0000    |
|     Waltz     |  0.4057 | 0.8000 |    0.3560    |    0.0283    |    0.0448    |
+---------------+---------+--------+--------------+--------------+--------------+

Q4~5: Tempo estimation using autocorrelation tempogram

In this task, we use autocorrelation tempogram to perform tempo estimation.

$ python3 Q4-5.py

+---------------+---------+--------+--------------+--------------+--------------+------------+------------+------------+
|     Genre     | P-score | ALOTC  | 1/2T P-score | 1/3T P-score | 1/4T P-score | 2T P-score | 3T P-score | 4T P-score |
+---------------+---------+--------+--------------+--------------+--------------+------------+------------+------------+
|   ChaChaCha   |  0.5064 | 0.9820 |    0.3512    |    0.0000    |    0.0000    |   0.1333   |   0.0000   |   0.0000   |
|      Jive     |  0.4544 | 0.9333 |    0.0000    |    0.0077    |    0.0000    |   0.5022   |   0.0000   |   0.0204   |
|   Quickstep   |  0.4331 | 0.8780 |    0.0000    |    0.0000    |    0.0000    |   0.4717   |   0.0114   |   0.0241   |
|     Rumba     |  0.4533 | 0.9184 |    0.4554    |    0.0000    |    0.0048    |   0.0099   |   0.0000   |   0.0000   |
|     Samba     |  0.3885 | 0.7674 |    0.4262    |    0.0000    |    0.0400    |   0.0210   |   0.0000   |   0.0000   |
|     Tango     |  0.5144 | 0.9535 |    0.2381    |    0.0000    |    0.0000    |   0.2021   |   0.0000   |   0.0000   |
| VienneseWaltz |  0.5071 | 0.9077 |    0.0445    |    0.0000    |    0.0000    |   0.2104   |   0.1377   |   0.0000   |
|     Waltz     |  0.2908 | 0.5364 |    0.3277    |    0.0313    |    0.0609    |   0.0000   |   0.0037   |   0.0000   |
+---------------+---------+--------+--------------+--------------+--------------+------------+------------+------------+

Q6: Tempo estimation by combine frequency and periodicity

This task the teacher ask us to find de wae to improve and outperform the above method.

How to improve?

Histogram of the most two probable tempo using Fourier tempogram:

Histogram of the most two probable tempo using autocorrelation tempogram:

As you can see, Fourier tempogram is easily to produce tempo which is multiple of the true tempo; on the other hand, autocorrelation tempogram is easily to produce tempo which is one half of the true tempo. So I decide to combine the two tempograms together, let each one supress others unwanted value and reserved the most probable one.

How to do?

And intuitive way to do it is mapping tempograms to the same domain, then multiply together. The implementation detailed is similar to this paper.

$ python3 Q6.py

+---------------+---------+--------+
|     Genre     | P-score | ALOTC  |
+---------------+---------+--------+
|   ChaChaCha   |  0.4962 | 0.9910 |
|      Jive     |  0.5229 | 0.9667 |
|   Quickstep   |  0.3957 | 0.8293 |
|     Rumba     |  0.4061 | 0.8776 |
|     Samba     |  0.3617 | 0.7326 |
|     Tango     |  0.6069 | 0.9884 |
| VienneseWaltz |  0.4444 | 0.8000 |
|     Waltz     |  0.4546 | 0.8091 |
+---------------+---------+--------+

The result shows that some genres have been improved, but some are not.

Q7: Beat tracking using dynamic programming

In this task, we also use the same dataset to perform beat tracking. The algorithm we used is describe here. I used CFP method to compute the tempo that can be used by the algorithm.

The program will output the precision, recall and f-score evaluate with tolerance of +-70 ms on each genre.

$ python3 Q7.py

+---------------+-----------+--------+----------+
|     Genre     | Precision | Recall | F-scores |
+---------------+-----------+--------+----------+
|   ChaChaCha   |   0.6381  | 0.9792 |  0.7727  |
|      Jive     |   0.9386  | 0.8072 |  0.8680  |
|   Quickstep   |   0.9262  | 0.6444 |  0.7600  |
|     Rumba     |   0.5558  | 0.9378 |  0.6979  |
|     Samba     |   0.4332  | 0.8356 |  0.5706  |
|     Tango     |   0.8807  | 0.9154 |  0.8977  |
| VienneseWaltz |   0.9132  | 0.6732 |  0.7750  |
|     Waltz     |   0.5526  | 0.8018 |  0.6543  |
+---------------+-----------+--------+----------+

Down beat tracking

This is a bonus question, and I just do some trial and error to see what will happen. I used the beats in Q7 to construct a bidirectional spectral flux novelty curve in beat level, and use the same algorithm in Q7 to find the path of downbeat with fix period of 4 samples (which means I assume the beats are isometric and have time signature of 4/4).

The result is apparently not good, but it's fun to try using traditional technique instead of fancy machine learning method.

$ python3 downbeat.py

+---------------+-----------+--------+----------+
|     Genre     | Precision | Recall | F-scores |
+---------------+-----------+--------+----------+
|   ChaChaCha   |   0.1204  | 0.1680 |  0.1403  |
|      Jive     |   0.0878  | 0.0679 |  0.0765  |
|   Quickstep   |   0.0823  | 0.0515 |  0.0634  |
|     Rumba     |   0.0673  | 0.1037 |  0.0816  |
|     Samba     |   0.1794  | 0.3167 |  0.2291  |
|     Tango     |   0.5912  | 0.5463 |  0.5679  |
| VienneseWaltz |   0.3285  | 0.1639 |  0.2187  |
|     Waltz     |   0.2155  | 0.2076 |  0.2114  |
+---------------+-----------+--------+----------+