summary of recent activity

The first 6 months…

The main activity over the first 6 months of the ShakeIt project addressed the following areas: i) to set up the project infrastructure, ii) to hold an internal workshop for all collaborators on the project, and iii) to design, conduct and report on the results of the first psychological experiment.

Development of Computational Models:

The main work conducted on the development of computational models was to re-implement existing code from a prior experiment on audio descriptors and the perception of groove (see Madison et al, 2011 in publications). This involved re-writing the code and adding comments to describe its operation. The code performs two types of important analysis relevant to the project: the detection of periodicity features (related to tempo and beat structure) and techniques for measuring microtiming deviations present in musical audio signals. The code is publicly available under GPL license at:

Groove Perception Experiment 1:

The aim of the first psychological experiment, PsychExpe1, was to explore the effect of microtiming deviations and groove. The entire process of running the experiment was conducted in several stages. First the experiment was designed during the internal workshop with Guy Madison (May, 2011), then a set of musical stimuli covering different musical styles and microtiming patterns and magnitudes were constructed. A piece of standalone software was created to the run the experiment allowing playback of the stimuli and the ability to recording ratings made by the participants. A set of 30 participants were recruited by emails sent to mailing lists at INESC and FEUP. Prior to taking the experiment, each participant was given clear instructions and signed a consent form to allow for their ratings to be analysed anonymously. Participants were paid 10€ for taking part. The experiment took place at INESC between 27th July and 12th August 2011.

Analysis of the experimental results demonstrated a largely negative effect for microtiming and groove, showing a clear tendency for groove ratings to decrease as a function of the magnitude of the microtiming. While we were confident about the design of the experiment and the validity of the musical stimuli created, the results raised the issue of whether musical training would be required to fully appreciate microtiming in music. To this end we plan to run a second psychological experiment using the same stimuli and design but using a group of expert participants with musical training.

The stimuli and scripts for applying microtiming deviations to MIDI files are available on here.

Conference calls

Here are lists of scientific conference calls and journal calls updated by our friends at MTG

Feel free to add some more

Effects of metrical levels on groove ratings and beat perception

This is in part a replication of the synchronisation to isochronous sequences with faster metrical level performed in the Spring of 2010. It comprised 9 levels of IOI (524-3126 ms) and 4 levels of metrical levels (1, 2, 3, and 4), that is 36 conditions (which means that the shortest interval in any sequence is 524/8 = 65 ms). For each condition, the participant had to synchronise 33 beats, which were indicated by a different sound.
For a perceptual experiment, the slowest and fastest tatum conditions are not that interesting, because the results are already known and also far from musically relevant. The tatum limit is around 300 ms (an educated guess; it might in fact be shorter still in the context of a rhythmical pattern, in which the event density is decreased by omitting some sounds or by loudness patterns downplaying the salience of some sounds).
So, for 1 ML we could omit 1600-3125 ms = 4 conditions. For the fastest levels we could omit tatums 65-300 ms = 12 conditions. That would reduce the experiment by > 1/3 from 36 to 20 conditions, but on the other hand destroy the possibility to perform linear stats such as ANOVA. Trials took 18-105 s, with a mean of 43 s, which means a total expt duration of 25 min. In this case, the sequence length should be long enough to ”feel the groove” and to hear at least a few of the longest cycles (3.125 s x 5? = 15-20 s). Then it takes time to rate and then it would be very nice to have data on how they perceive the beat, that is, ask them to indicate that with some 5-10 responses, and how that might relate to the groove, in particular in these quite ambiguous patterns. So that could also take up to 20 s, which means we end up approx the same 25 min.
It would be very good to be able to compare this with the more ecological condition in which each ML is conveyed by a different sound rather than just loudness, i.e., is clearly distinguishable. All of the trials with faster beats could be shortened, however, in the response part, so we could probably cut another 5-7 minutes. Doubling the expt with a ”percussion band” version would give a total of about 40 minutes then, which is acceptable.
These stimulus sequences can quite easily be generated by my MLP program. That program can also record beats, but does not feature any rating capability. So the question is how to get the best of both worlds. I suggest we do ratings and beat beating in another application with audio files.
Would there be a point to choose a constant for the geometrical increase in IOI that aligns IOIs across metrical levels?
Predictions would be that for only loudness-separated ML:
1. Each ML increases groove up to tatum ~300 ms
2. Beat tempo will be narrow around 100-120 BPM
3. Groove is highest for optimal tempi, everything else being the same, but this effect is much smaller than the effect of ML
Predictions for instrument-sound-separated ML:
1. Each ML increases groove up to tatum < 300 ms
2. Beat tempo will be more dispersed around 100-120 BPM than for loudness-separated ML
3. Beat tempo will unrelated or at least less strongly related to groove than for loudness-separated ML, because beat tempo will be much slower for the slow multi-level sequences, although their groove will be determined by ML.
It seems that the ”sparse fast metrical levels” experiment would benefit from the knowlegde provided by the above design, so I suggest we take that as a second step. For example, if we see no increased groove even for medium tatums, then there’s no point in sparsing out faster tatums.