Jean-Louis DURRIEU
PhD Candidate at the Ecole Nationale Supérieure des Télécommunications (ENST or TELECOM ParisTech)
PhD Candidate at the Ecole Nationale Supérieure des Télécommunications (ENST or TELECOM ParisTech)
Below
are some resulting sounds from the source separation algorithm we have
designed, based on a previous melody detection (in Durrieu et al.,
SINGER MELODY EXTRACTION IN POLYPHONIC SIGNALS USING SOURCE SEPARATION
METHODS, see the publications
page), followed by a spectral
Wiener filtering of the original signal. The songs we used are from the
ISMIR 2004 Audio Melody
Extraction Contest database (see the dedicated ismir 2004 website for more
details and
to download the test set and the reference files). We suggest you to
use headphones in order to be able to hear slight differences and
artifacts in the resulting sounds. If you experience any problems with
the flash player, you can also check this alternate page with
direct links to the files.
The parameters for the analysis of MIREX 2004 database songs are:
- sampling rate: 44100Hz
- length of analysis windows: 46.44 ms (2048 samples)
- hopsize: 5.8 ms (256 samples)
- frequency range for melody detection: Fmin = 100Hz, Fmax = 800Hz (except when precised)
- frequency quantization for the melody: discretized to every 8th tone (that is to say there are 48 frequencies per octave)
- length of the songs: between 14s and 25s
- Title: title name of the song in original test set
- Original: the original song
- Sep. Singer: estimated/separated singer signal
- Sep. Music: estimated/separated music signal
- Remix: left channel = estimated singer, right channel = estimated music
- PitchMatch: percentage of "correctness" in pitch estimation in the song's pitched frames (see the ismir 2004 website for a description)
- TotalMatch: percentage of "correctness" in pitch estimation in the whole song (also taking into account the silences in the singer reference track)
| Vx Recall | Vx False Alm | Vx d' | Raw pitch | Raw Chroma | Overall Acc | |
| dressler | 89.9% | 23.5% | 2.00 | 80.0% | 82.9% | 77.3% |
| ryynanen | 80.4% | 15.4% | 1.88 | 75.5% | 78.2% | 72.1% |
| poliner | 90.2% | 35.7% | 1.66 | 72.6% | 75.7% | 69.3% |
| sutton | 67.6% | 17.0% | 1.41 | 59.1% | 62.5% | 55.7% |
| brossier | 99.5% | 97.1% | 0.66 | 48.3% | 61.7% | 39.8% |
| durrieu | 99.99% | 99.81% | 0.72 | 76.13% | 80.02% | 61.92% |
| Title |
Original |
Sep.
Singer |
Sep. Music |
Remix |
Pitch Match (%) |
Total Match (%) |
| opera_fem2 |
70.2 |
63.7 |
||||
| opera_fem4 |
78.8 |
80.2 |
||||
| opera_male3 |
65.5 |
66.8 |
||||
| opera_male5 |
82.8 |
77.7 |
||||
| daisy1 |
85.5 |
72.1 |
||||
| daisy2 |
85.3 |
74.6 |
||||
| daisy3 |
84.7 |
84.7 |
||||
| daisy4 |
90.5 |
90.5 |
||||
| pop1 |
74.2 |
62.7 |
||||
| pop2 |
79.8 |
63.8 |
||||
| pop3 |
76.8 |
62.1 |
||||
| pop4 |
79.4 |
64.8 |
||||
| jazz1 |
73.4 |
71.1 |
||||
| jazz2 |
71.2 |
67.3 |
||||
| jazz3 |
78.8 |
52.3 |
||||
| jazz4 |
71.4 |
57.0 |
||||
| midi1 |
74.6 |
70.7 |
||||
| midi2 |
81.9 |
81.9 |
||||
| midi3 |
60.0 |
60.5 |
||||
| midi4 (Fmax=1200) |
84.8 |
73.5 |
Some results on a database I got from http://www.ee.columbia.edu/~graham/mirex_melody/, which seems to be the files "competitors" for the MIREX 2005 Melody Extraction Task could use to tune their algorithms.
The parameters are almost the same as before, except that the hopsize for the analysis windows is equal to 10ms (441 samples) to fit the given groundtruth.
| Title |
Original |
Sep.
Singer |
Sep. Music |
Remix |
Pitch Match (%) |
Total Match (%) |
| train01 |
80.9 | 55.1 | ||||
| train02 |
56.9 | 37.6 | ||||
| train03 |
77.7 | 48.3 | ||||
| train04 |
71.2 | 61.5 | ||||
| train05 |
76.5 | 60.0 | ||||
| train06 |
58.7 | 29.3 | ||||
| train07 |
72.2 | 56.4 | ||||
| train08 |
78.7 | 58.9 | ||||
| train09 |
85.1 |
70.4 |
... and below another example showing the possible use of our model for other instruments, in excerpts of take five (being a saxophonist, this example was compulsory for me!):
| Title |
Original |
Sep.
Saxophone |
Sep.
Background Music |
Remix |
| takefive01 | ||||
| takefive02 |
