2012:Audio Chord Estimation
From MIREX Wiki
Evaluation of Chord Transcriptions
Before the final description of the chord evaluation goes live here, please see the discussion based on the The Utrecht Agreement on Chord Evaluation.
This task requires participants to extract or transcribe a sequence of chords from an audio music recording. For many applications in music information retrieval, extracting the harmonic structure of an audio track is very desirable, for example for segmenting pieces into characteristic segments, for finding similar pieces, or for semantic analysis of music.
The extraction of the harmonic structure requires the detection of as many chords as possible in a piece. That includes the characterisation of chords with a key and type as well as a chronological order with onset and duration of the chords.
Although some publications are available on this topic [1,2,3,4,5], comparison of the results is difficult, because different measures are used to assess the performance. To overcome this problem an accurately defined methodology is needed. This includes a repertory of the findable chords, a defined test set along with ground truth and unambiguous calculation rules to measure the performance.
Three datasets are used to evaluate chord transcription accuracy:
Christopher Harte`s Beatles dataset consisting of annotations of 12 Beatles albums.
The text annotation procedure of musical chords that was used to produce this dataset is presented in .
Queen and Zweieck dataset
Matthias Mauch's Queen and Zweieck dataset consisting of 38 songs from Queen and Zweieck.
Billboard dataset (abridged)
An abridged version of Ashley Burgoyne's Billboard dataset , consisting of about 200 songs for training (previously published) and 200 songs for testing (to be published for the first time at ISMIR).
Example ground-truth file
The ground-truth files take the form:
... 41.2631021 44.2456460 B 44.2456460 45.7201230 E 45.7201230 47.2061900 E:7/3 47.2061900 48.6922670 A 48.6922670 50.1551240 A:min/b3 ...
The segmentation score will be calculated using directional hamming distance as described in . An over-segmentation value (m) and an under-segmentation value (f) will be calculated and the final segmentation score will be calculated using the worst case from these two i.e:
segmentation score = 1 - max(m,f)
m and f are not independent of each other so combining them this way ensures that a good score in one does not hide a bad score in the other. The combined segmentation score can take values between 0 and 1 with 0 being the worst and 1 being the best result.-- Chrish 17:05, 9 September 2009 (UTC)
For recall evaluation, we may define a different chord dictionary for each level of evaluation (dyads, triads, tetrads etc). Each dictionary is a text file containing chord shorthands / interval lists of the chords that will be considered in that evaluation. The following dictionaries are proposed:
For dyad comparison of major/minor chords only:
For comparison of standard triad chords:
For comparison of tetrad (quad) chords (currently only for the Beatles and Queen and Zweieck datasets):
For each evaluation level, the ground truth annotation is compared against the dictionary. Any chord label not belonging to the current dictionary will be replaced with an "X" in a local copy of the annotation and will not be included in the recall calculation.
Note that the level of comparison in terms of intervals can be varied. For example, in a triad evaluation we can consider the first three component intervals in the chord so that a major (1,3,5) and a major7 (1,3,5,7) will be considered the same chord. For a tetrad (quad) evaluation, we would consider the first 4 intervals so major and major7 would then be considered to be different chords.
For the maj/min evaluation (using the first example dictionary), using an interval comparison of 2 (dyad) will compare only the first two intervals of each chord label. This would map augmented and diminished chords to major and minor respectively (and any other symbols that had a major 3rd or minor 3rd as their first interval). Using an interval comparison of 3 with the same dictionary would keep only those chords that have major and minor triads as their first 3 intervals so augmented and diminished chords would be removed from the evaluation.
After the annotation has been "filtered" using a given dictionary, it can be compared against the machine generated estimates output by the algorithm under test. The chord sequences described in the annotation and estimate text files are sampled at a given frame rate (in this case 10ms per frame) to give two sequences of chord frames which may be compared directly with each other. For calculating a hit or a miss, the chord labels from the current frame in each sequence will be compared. Chord comparison is done by converting each chord label into an ordered list of pitch classes then comparing the two lists element by element. If the lists match to the required number of intervals then a hit is recorded, otherwise the estimate is considered a miss. It should be noted that, by converting to pitch classes in the comparison, this evaluation ignores enharmonic pitch and interval spellings so the following chords (slightly silly example just for illustration) will all evaluate as identical:
C:maj = Dbb:maj = C#:(b1,b3,#4)
Basic recall calculation algorithm:
1) filter annotated transcription using chord dictionary for a defined number of intervals
2) sample annotated transcription and machine estimated transcription at 10ms intervals to create a sequence of annotation frames and estimate frames
3) start at the first frame
4) get chord label for current annotation frame and estimate frame
5) check annotation label:
IF symbol is 'X' (i.e. non-dictionary)
THEN ignore frame (record number of ignored frames)
ELSE compare annotated/estimated chords for the predefined number of intervals
increment hit count if chords match
6) increment frame count
7) go back to 4 until final chord frame --Chrish 17:05, 9 September 2009 (UTC)
Audio tracks will be encoded as 44.1 kHz 16bit mono WAV files.
The expected output chord transcription file for participating algorithms is that proposed by Christopher Harte .
Hence, algorithms should output text files with a similar format to that used in the ground truth transcriptions. That is to say, they should be flat text files with chord segment labels and times arranged thus:
start_time end_time chord_label
with elements separated by white spaces, times given in seconds, chord labels corresponding to the syntax described in  and one chord segment per line.
The chord root is given as a natural (A|B|C|D|E|F|G) followed by optional sharp or flat modifiers (#|b). For the evaluation process we may assume enharmonic equivalence for chord roots. For a given chord type on root X, the chord labels can be given as a list of intervals or as a shorthand notation as shown in the following table:
|major||X:(1,3,5)||X or X:maj|
|possible 6th triad:|
|minor 7/b5 (ambiguous - could be either of the following)|
|minor-minor7-b5 (a half diminished-7th)||X:(1,b3,b5,b7)||X:hdim7|
|omitted from list on wiki:|
Please note that two things have changed in the syntax since it was originally described in . The first change is that the root is no longer implied as a voiced element of a chord so a C major chord (notes C, E and G) should be written C:(1,3,5) instead of just C:(3,5) if using the interval list representation. As before, the labels C and C:maj are equivalent to C:(1,3,5). The second change is that the shorthand label "sus2" (intervals 1,2,5) has been added to the available shorthand list.--Chrish 17:05, 9 September 2009 (UTC)
We still accept participants who would only like to be evaluated on major/minor chords and want to use the number format which is an integer chord id on range 0-24, where values 0-11 denote the C major, C# major, ..., B major and 12-23 denote the C minor, C# minor, ..., B minor and 24 denotes silence or no-chord segments. Please note that the format is still the same
start_time end_time chord_number
Systems are supposed to print out the onset-offset times as opposed to MIREX 2008 chord output format where only onset were used.
Command line calling format
Submissions have to conform to the specified format below:
extractFeaturesAndTrain "/path/to/trainFileList.txt" "/path/to/scratch/dir"
Where fileList.txt has the paths to each wav file. The features extracted on this stage can be stored under "/path/to/scratch/dir" The ground truth files for the supervised learning will be in the same path with a ".txt" extension at the end. For example for "/path/to/trainFile1.wav", there will be a corresponding ground truth file called "/path/to/trainFile1.wav.txt" .
doChordID.sh "/path/to/testFileList.txt" "/path/to/scratch/dir" "/path/to/results/dir"
If there is no training, you can ignore the second argument here. In the results directory, there should be one file for each testfile with same name as the test file + .txt .
Programs can use their working directory if they need to keep temporary cache files or internal debuggin info. Stdout and stderr will be logged.
All submissions should be statically linked to all libraries (the presence of dynamically linked libraries cannot be guaranteed).
All submissions should include a README file including the following information:
- Command line calling format for all executables and an example formatted set of commands
- Number of threads/cores used or whether this should be specified on the command line
- Expected memory footprint
- Expected runtime
- Any required environments (and versions), e.g. python, java, bash, matlab.
Time and hardware limits
Due to the potentially high number of particpants in this and other audio tasks, hard limits on the runtime of submissions are specified.
A hard limit of 24 hours will be imposed on runs (total feature extraction and querying times). Submissions that exceed this runtime may not receive a result.
Please write your comments below with your name and date.
Somewhere in the email discussion on the MIREX list, there was a mention that the recent systems run on the Beatles/Queen/Zweieck dataset might have over-learnt the properties of this dataset. I just wondered whether, during or post-MIREX, there was any way to formally/experimentally demonstrate this? I mean, beyond making the observation that there is a "drop" in performance from an open dataset to a closed one. The issue would seem particularly pertinent with regard to this dataset since it's been public for sometime. (Matthew Davies, 9th August)
name / email
1. Harte, C.A. and Sandler, M.B. (2005). Automatic chord identification using a quantised chromagram. Proceedings of 118th Audio Engineering Society's Convention.
2. Sailer, C. and Rosenbauer K. (2006). A bottom-up approach to chord detection. Proceedings of International Computer Music Conference 2006.
3. Shenoy, A. and Wang, Y. (2005). Key, chord, and rythm tracking of popular music recordings. Computer Music Journal 29(3), 75-86.
4. Sheh, A. and Ellis, D.P.W. (2003). Chord segmentation and recognition using em-trained hidden markov models. Proceedings of 4th International Conference on Music Information Retrieval.
5. Yoshioka, T. et al. (2004). Automatic Chord Transcription with concurrent recognition of chord symbols and boundaries. Proceedings of 5th International Conference on Music Information Retrieval.
6. Harte, C. et al. (2005). Symbolic representation of musical chords: a proposed syntax for text annotations. Proceedings of 6th International Conference on Music Information Retrieval.
7. Papadopoulos, H. and Peeters, G. (2007). Large-scale study of chord estimation algorithms based on chroma representation and HMM. Proceedings of 5th International Conference on Content-Based Multimedia Indexing.
8. Abdallah, S. et al. (2005). Theory and Evaluation of a Bayesian Music Structure Extractor (pp. 420-425) Proc. 6th International Conference on Music Information Retrieval, ISMIR 2005.
9. John Ashley Burgoyne et al. (2011). An expert ground-truth set for audio chord recognition and music analysis (pp. 633–638) Proc. 12th International Society for Music Information Retrieval Conference, ISMIR 2006. (PDF)