i.e. every 4 beats (one step in the time variable t) the model progresses to a new Xt hidden state. Each X produces a probability distribution for a progression along Ys hidden states, such that a time step in the variable s is 1/4 of a step in t. Each Y has an emission distribution to observable states Zs (which also occur on each beat.
This is assuming a 4/4 time signature but could also be set to 3/4 or others by adjusting the time step ratios.
Ideally, an even longer time interval could be created for 4 measure sequences, making up a “verse” since the larger-scale changes in music are noticeable at this scale
Does anyone know if this has been explored or have any thoughts?