The homepage of the "A Song for Bayes" arrangment

The homepage of the choir vox11 including music samples


Music experience modelling, March 9th 2011

One of the unique properties of music as art compared to other arts is that the performance and therefore consumption of musics takes places in time. Music is basically a sequence of moments with sounds. 

A sound(valid) model of music experience should reflect this fact. 

The bayesian models can integrate the time property in markov chains by duplicating the model and describe the relation between the timesteps. 


Above we see a bayesian model of Musictherapy based that consists of 3 timesteps, t1,t2 and t3. Timesteps that can be sequences of music experience. 

Each timestep is modelled the same way: "Music" causes a certain "Arousal" which again causes a changed sensor measurement in terms of 2 sensor measurements: HRV(Heart Rate Variability) and GSR(Galvanic Skin Response).


Due to the fact that arousal in one timestep(situation) influences what can happen next the Arousal in step t1 has an impact(causes) on the Arousal in step t2. Eg. if the music experience in t1 gives the musictherapy client a strong positive feeling, eg being happy, it is unlikely that the Arousal in t2 is a very negative feelling, eg. being very sad. Not to say that it is impossible, but the probability for such case is low. 

And the connection/causes/impacts in bayesian nets are not deterministic, but probabilitic. Regardless of the fact that the bayesian net is a markov chain net or not.  
And the impact is evidence that the model should capture to be a more accurate model. 

The weak point in the model is that the nodes "Arousal" and "Music" are fluffy-fluffy terms. Not aristolian terms. The soundness/validity of the such models depends heavily on the model designers ablitity to integrate the scientifically  proved terms in the model .