2009:Audio Melody Extraction
- 1 Description
- 2 Discussions for 2009
- 2.1 New evaluations for 2009?
- 2.2 Chao-Ling's Comments 14/07/2009
- 2.3 Vishu's Comments 20/07/2009
- 2.4 Chao-Ling's Comments 21/07/2009
- 2.5 Chao-Ling's Comments 22/07/2009
- 2.6 About MIR-1K - Jean-Louis, 26/07/2009
- 2.7 Chao-Ling's Comments 27/07/2009
- 2.8 Andreas Ehmann's Comments 12/08/2009
- 2.9 Jean-Louis' Comments 17/08/2009
- 2.10 Andreas Ehmann's Comments 18/08/2009
- 2.11 Vishu's comments 19/08/2009
- 2.12 Chao-Ling's Comments 02/09/2009
- 2.13 Vishu's comments 03/09/2009
- 2.14 Chao-Ling's Comments 03/09/2009
- 2.15 Vishu's comments 03/09/2009
- 2.16 Morten's comments 05/09/2009
- 2.17 Jean-Louis' comments 06/09/2009
- 2.18 Chao-Ling's Comments 07/09/2009
- 2.19 Vishu's Comments 07/09/2009
- 2.20 Jean-Louis' Comments 08/09/2009
- 2.21 Chao-Ling's Comments 08/09/2009
- 2.22 Sihyun-Joo's Comments 08/09/2009
- 2.23 Vishu's Comments 11/09/2009
- 3 Dataset
- 4 Output Format
- 5 Relevant Development Collections
- 6 Potential Participants
- 7 Results
The aim of the MIREX audio melody extraction evaluation is to identify the melody pitch contour from polyphonic musical audio.
The task consists of two parts:
- Voicing detection (deciding whether a particular time frame contains a "melody pitch" or not),
- pitch detection (deciding the most likely melody pitch for each time frame).
We structure the submission to allow these parts to be done independently, i.e. it is possible (via a negative pitch value) to guess a pitch even for frames that were being judged unvoiced. Algorithms which don't perform a discrimination between melodic and non-melodic parts are also welcome!
Discussions for 2009
Your comments here.
New evaluations for 2009?
We would like to know if there would be potential participants for this year's evaluation on Audio Melody Extraction.
There has also been an interest last year in evaluating the results at note levels (and not at a frame by frame level), following the multipitch evaluation. However, it has not been done, probably because of both a lack of participants and of database. Would there be more people this year?
cheers, Jean-Louis, 9th July 2009
Chao-Ling's Comments 14/07/2009
Hi everyone. I would like to suggest that we have a separate evaluation on the songs where the main melody is carried by the human singing voice as opposed to other musical instruments (like Vishu's comment in MIREX2008). We proposed a pitch extraction approach for singing voices and may not be likely to perform well for other instruments.
In addition, we have prepared a dataset called MIR-1K and would like to add it as part of the training/evaluation dataset. It contains 1000 song clips recorded at 16 kHz sample rate with 16-bit resolution. The duration of each clip ranges from 4 to 13 seconds, and the total length of the dataset is 133 minutes. These clips were extracted from 110 karaoke songs which contain a mixed track and a music accompaniment track.
Vishu's Comments 20/07/2009
Chao-Ling, your dataset sounds exciting. I think the community would benefit greatly from the addition of such a large database. Last year we too had contributed some data (Indian classical music). Here are some points with respect to your current dataset to make it conform to previous data formats.
- We need to have mixed (voice + accompaniment) mono tracks, so if you would like to mix it yourself (based on some SNR considerations) or if you think just giving equal weight to left and right channels of your files is acceptable, let us know which.
- I see that your ground-truth's (.pv files) are in semitones. Since all the previous reference pitch values are in Hz either you could convert them to Hz and pass them on or pass on the exact conversion formula (Hz to cents) you have used to the evaluators (IMMERSEL).
- Finally, the previous ground-truth pitch files were in two column format (TimeStamp_sec PitchValue_Hz) available every 10 ms. Your files are single-column. Please let us know at exactly which time-instant is your first window centered and the conversion can be done accordingly.
Thanks again for your efforts.
Chao-Ling's Comments 21/07/2009
Hi Vishu and everybody! Thank you for your suggestions. My responses are as follows:
- The left and right channels were adjusted to have equal weight. I prefer to provide both channels because they are good for evaluating algorithms at different SNR settings.
- I will provide pitch ground-truth in Hz.
- The first window is centered in 20ms (the window size is 40ms and the overlap is 20ms). Note that the last window is discarded if its time is less than 40ms, so some .pv files might have one less point.
I will provide new .pv files that have both ground-truth in Hz and time stamp column. I will also make this dataset smaller that discards the parts that are unrelated to this task.
Chao-Ling's Comments 22/07/2009
Okay guys, here is the new files of the dataset: MIR-1K for MIREX. Plesea feel free to let me know if there are problems.
About MIR-1K - Jean-Louis, 26/07/2009
Such a big database indeed is good news for the relevance of the evaluation. However, in the spirit of MIREX (if any), it may have been good to keep some part of it "hidden" to the participants, so as to perform evaluation on a test database, on which no one could have tuned their algorithms.
Do you, by any chance, have another 1000 songs that we could use for that purpose? :p Well, otherwise, that still makes it a good database for evaluation and comparison.
By the way, I have trouble checking the above mentioned rar archive file: on my ubuntu, it says the archive type is not supported. Any idea?
One last thing: is anyone interested in evaluating note-wise transcription (as in the multi-f0 evaluation task)? If so, is there any annotation of that type for MIR-1K?
Chao-Ling's Comments 27/07/2009
Hi Jean-Louis and everyone!
Unfortunately, I donΓÇÖt have another dataset. Even if I do, it is not ΓÇ£hiddenΓÇ¥ from me :S.
The rar file can be extracted in ubuntu by this program: winrar for Linux. However, MIR-1K does not contain the annotation for evaluating note-wise transcription.
Andreas Ehmann's Comments 12/08/2009
Hi guys! We are quietly ramping up for this year's MIREX. The dataset is quite exciting. Although it's not 'hidden' I think it's more than useable. ADC04 isn't quite withheld either. I guess my main concern from a logistics point of view is that it is pretty big! Some of the melody algorithms are on the slow side, so crunching through that many minutes of audio might tie up our machines quite a bit. So I guess our options are try and make sure the algorithms are fast enough, or we can maybe choose a subset of the 1000 to evaluate against.
Jean-Louis' Comments 17/08/2009
Hi everyone, It feels like andreas' comment on the speed of some algorithms was sort of referring to Pablo's and my program from last year. I can't however guarantee that this year's algorithms will be any faster...
I guess working on subsets of the database is a good option. Maybe a few 100 snippets from it. Chao-Ling: is the database homogeneous, such that one can grab randomly any excerpt and get a representative dataset, or is there a smart way of choosing these excerpts?
Andreas Ehmann's Comments 18/08/2009
I think I am going to sample rate convert the MIR-1K dataset to 44.1kHz (from the 16kHz it is now). Naturally we will have dead space in the spectrum, but I am already envisioning systems having hard coded (in samples) frame lengths and hops. Sound reasonable? That or everyone has to ensure they are robust to multiple SR's. It's easy to do on my end though, and that way everything will be 44.1kHz.
Vishu's comments 19/08/2009
Andreas, wrt to our specific entry/entries, they will be SR independent. So 16 kHz or 44.1 kHz doesn't really matter. In the interest of data homogeneity however, maybe 44.1 kHz is preferable.
I have a question regarding deadlines. Are we following the Sept. 8 deadline, as posted on the MIREX 2009 homepage, or do you think we could push this up a bit. I ask this because we intend to submit two algorithms this year and the second one may not be ready by Sept 8.
Chao-Ling's Comments 02/09/2009
Hi Jean-Louis and all,
Sorry for my late reply. I worked very hard to build a "hidden" dataset for this task. The length of the dataset is around 167 minutes (374 clips with length 20~40 secs each). The dataset was built in the same way as MIR-1K with the same format (16kHz,16bits). I would like to know how do we evaluate our algorithms with this dataset and how do I provide it to the committee?
Vishu's comments 03/09/2009
Chao-Ling: Last year we had contributed an Indian classical music dataset. I had corresponding with Mert Bay (email@example.com) at that time.
Chao-Ling's Comments 03/09/2009
Thx Vishu. Andreas Ehmann will set up a dropbox account for me to upload the dataset. Besides, I would like to know what SNR value should be used to mix the singing voice and accompaniment for the evaluation. Any suggestion?
Vishu's comments 03/09/2009
Chao-Ling: From my point of view, it would be useful to divide the dataset into two halves. One with an audibly acceptable SNR (between 5 and 10 dB) and the other with a lower, and therefore tougher, SNR (0 dB). Of course, a lot also depends on the nature of the accompaniment i.e. lower SNR on simple (solo) accompaniment, like a single flute, may not provide as much of a challenge as a relatively higher SNR but with more complex accompaniment, like rock music. Note that I use the terms 'simple' and 'complex' purely from a point of view of signal complexity. Since you are most familiar with the data, you could divide it as you see fit.
Morten's comments 05/09/2009
Chao-Ling and Vishu: I think that it would be preferable if a part of the dataset is mixed both at an "audibly acceptable" SNR and at a "toughter" SNR. It would be difficult to conclude anything if the two mixing levels was used on two different datasets - does the results then depend on the mixing level or on the dataset?
Jean-Louis' comments 06/09/2009
I agree with Morten. If we want to compare the results with varying mixing level, we should evaluate the algorithms on the same songs, with different "SNR"s (if one can indeed consider the background music as "noise" :-) ).
By the way, any possible deadline extension? It is going to be a bit tight for us, I'm afraid. But we try our best, because we might also provide the slowest programs. We should probably not delay that too much :)
Oh, I was also thinking, in the article by Poliner /et al/, "Melody Transcription From Music Audio: Approaches and Evaluation." (IEEE Transactions on Audio, Speech & Language Processing 15(4): 1247-1256 (2007)), they managed to gather information concerning the error distribution of the algorithms, i.e. the histogram of differences (in note on the Western musical scale) between the estimated pitches and the ground-truth. Would you think, at IMIRSEL, that it would be possible to provide that sort of statistics after the evaluation?
Also, I'd like some practical information about the databases: with the former databases (ADC04 and MIREX05), the fundamental frequency range is around 60Hz to 1000Hz. Note that there were some errors in my version of the annotations for MIREX05 (the minimum value for f0 was something like 8Hz). What about MIREX08 and MIR-1K? I believe MIR-1K is going from 80 to 700Hz. Should we expect something else from the newly contributed data?
At last, we'd like to thank Chao-Ling for his efforts in building this hidden dataset! Can you give some details on it? Are they the same singers? The same songs?
Chao-Ling's Comments 07/09/2009
I agree with Morten and Jean-Louis about the evaluation at different SNRs. By the way, to show the histogram of differences is a good idea. It can let us know more about the performance of each algorithm. About the new dataset, I am thinking about releasing it after the evaluation. It is good for the researchers in this field but might be disadvantageous for the MIREX in the future unless someone can provide a new dataset. Any suggestion?
Here are some details about the new dataset: Singers: 15 people->6 females, 9males (corpus size from females and males are more or less the same). Only one of the female singers is included in MIR-1K, and she contributed about 2% of the new dataset. (If you guys prefer a completely singer independent dataset, we can remove this part.)
Song numbers: 167 minutes in length, comprised of 374 clips from 102 songs, with 20 to 40 seconds in each clip. 15 songs of them also appear in MIR-1K.
Fundamental frequency range: 46Hz to 784Hz, 99.95% of them are in the range of 80Hz to 640Hz.
Vishu's Comments 07/09/2009
Different SNRs on the same dataset definitely would be the way to go - Thanks Morten and JL.
Details of the MIREX08 dataset: The data consists of four 1-min long clips from different segments of 2 north Indian classical vocal performances (1 male and 1 female performer), so that makes a total of eight minutes of audio. Typically a single performance lasts for 20-25 min., with singing and instrumental playing style variations over the course of the performance. Musical sound sources present in the recordings:
- Singing voice
- Perpetually present drone. This is an Indian instrument called the tanpura.
- Secondary melodic instrument : This is an instrument called the harmonium similar to the accordion.
- Tonal percussion: These are a pair of Indian drums called Tablas capable of producing pitched sounds.
So there may be a maximum of four pitches and a minimum of one pitch present at any given time. F0 range is from 100 to 600 Hz.
Finally, taking a cue from JL and encouraged by Stephen Downie's most recent email, I propose we extend the deadline to 15th September. Is that acceptable?
Jean-Louis' Comments 08/09/2009
It would definitely be hard for us to provide our systems today... As soon as we have something, we will upload it. Could we push September 15th as a new deadline? Which date would be a hard deadline after which you could not (at IMIRSEL) process the programs anymore?
Concerning the task itself, and especially the new dataset: after listening to some of the excerpts and after some tests, it felt that this data is suited for "singing voice f0 estimation", especially using the provided "SNR" (at 0dB). For some reason, it is difficult, even for a listener to focus on the singer. Maybe the general balance of the singing voice is not only depending on the "SNR" itself? Of course, if you design your algorithm in order to specifically estimate the singing voice (with a classification scheme at one stage), then you might perform better on this database. There are some excerpts in MIR-1K where, for instance, the singing voice is backed by some flute at one or two octaves higher. The balance between these two contributions is quite even, and one could argue the fact that the voice is indeed intended to be the "lead melody instrument". I guess one could say that there is an indeterminacy, there... which is addressed only if you consider that the task is to transcribe the _vocal_ part.
Chao-Ling, could you confirm whether our understanding of this new dataset is correct? What were the purpose and the assumptions that led to the development of your database? I think it is important to understand these well, especially when analyzing the results of our algorithms on such a database.
By the way, I guess for the previous datasets, we also had a few examples on which the concept of "main melody" was not quite well defined... Historically, it seems that the ADC04 was allowing a rather wide definition of "leading melody" (including other instruments such as the saxophone as leading instrument), while the following evaluations were focusing more on popular songs, expecting the algorithms to extract the singing voice instead (without explicit notice, though).
I guess for forth-coming evaluations at MIREX, "Audio Melody Extraction" should be rephrased, or at least the assumptions on the melody to be extracted should be clearer. That may need some brain-storming, so I hope we can talk about it in Kobe, together :-D
For the time being, let us have the evaluation done the old way. I guess there is not enough time left to change everything right now!
P.S.: I added descriptions of the new dataset and MIREX08 dataset in the "dataset" and "relevant development set" sections. I renamed the latter, from "test" to "development", since this term is better suited for the content of the section... Please feel free to correct or extend this section!
Chao-Ling's Comments 08/09/2009
Hi JL and everyone,
We developed the dataset according to the article by Poliner et al.,"Melody Transcription from Music Audio: Approaches and Evaluation." (IEEE Transactions on Audio, Speech & Language Processing 15(4): 1247-1256 (2007)) which defines that the lead vocal constitutes the melody. However, I agree with JL that the definition is not so clear in some cases. Maybe ΓÇ£singing voice f0 estimationΓÇ¥ is more appropriate. This is also the reason that I proposed to have a subtask evaluation for vocal songs earlier. Just like what JL said, we should talk about it here and in Kobe together :).
Sihyun-Joo's Comments 08/09/2009
It's already due date. Do you finish your works well? I just left a message to ask about the due date. Actually, I want to postpone the due date for a week or a couple of days just like JL. I sent e-mail to request a new deadline, I haven't got the reply yet. However, as I mentioned before, it's already 8th September. So, if anybody knows whether it is possible to change the deadline, please let me know it by answer comment or e-mail (firstname.lastname@example.org).
Vishu's Comments 11/09/2009
Chao-Ling and Andreas: I have one (maybe redundant) concern about the MIR-1k data-set. I noticed that in the ground-truth (.pv) files, the first time-stamp is at 0.01 sec. For the ground-truth files of the previous data-sets, the first time-stamp was always at 0.0 sec. I hope this does not affect the evaluation-code since even a 1-frame offset in comparison can lead to significantly different accuracies. Or maybe a dummy pitch value could be inserted at 0.0 sec in the MIR-1k ground-truth files.
- MIR-1K database : dataset for Mirex, Karaoke recordings of Chinese songs. Instruments: singing voice (male, female), synthetic accompaniment.
- MIREX08 database : 4 excerpts of 1 min. from "north Indian classical vocal performances", instruments: singing voice (male, female), tanpura (Indian instrument, perpetual background drone), harmonium (secondary melodic instrument) and tablas (pitched percussions).
- MIREX05 database : 25 phrase excerpts of 10-40 sec from the following genres: Rock, R&B, Pop, Jazz, Solo classical piano.
- ISMIR04 database : 20 excerpts of about 20s each.
- CD-quality (PCM, 16-bit, 44100 Hz)
- single channel (mono)
- manually annotated reference data (10 ms time grid)
- In order to allow for generalization among potential approaches (i.e. frame size, hop size, etc), submitted algorithms should output pitch estimates, in Hz, at discrete instants in time
- so the output file successively contains the time stamp [space or tab] the corresponding frequency value [new line]
- the time grid of the reference file is 10 ms, yet the submission may use a different time grid as output (for example 5.8 ms)
- Instants which are identified unvoiced (there is no dominant melody) can either be scored as 0 Hz or as a negative pitch value. If negative pitch values are given the statistics for Raw Pitch Accuracy and Raw Chroma Accuracy may be improved.
Relevant Development Collections
- Graham's collection: you find the test set here and further explanations on the pages http://www.ee.columbia.edu/~graham/mirex_melody/ and http://labrosa.ee.columbia.edu/projects/melody/
- For the ISMIR 2004 Audio Description Contest, the Music Technology Group of the Pompeu Fabra University assembled a diverse of audio segments and corresponding melody transcriptions including audio excerpts from such genres as Rock, R&B, Pop, Jazz, Opera, and MIDI. (full test set with the reference transcriptions (28.6 MB))
- Vishweshwara Rao & Preeti Rao (Indian Institute of Technology Bombay, India)
- Jean-Louis Durrieu, Ga├½l Richard and Bertrand David (Institut T├⌐l├⌐com, T├⌐l├⌐com ParisTech, CNRS LTCI, Paris, France): 2 systems submitted (2009/09/19)
- Chao-Ling Leon Hsu, Jyh-Shing Roger Jang, and Liang-Yu Davidson Chen (Department of Computer Science, National Tsing-Hua University, Hsinchu, Taiwan)
- Morten Wendelboe (Institute of Computer Science, Copenhagen University, Denmark)
- Sihyun Joo & Seokhwan Jo & Chang D. Yoo (Korea Advanced Institute of Science and Technology, Daejeon, Korea)
- Pablo Cancela (email@example.com) Montevideo, Uruguay
- Hideyuku Tachibana, Takuma Ono, Nobutaka Ono, and Shigeki Sagayama (The University of Tokyo, Japan)