Consultation des notices

A classical problem with the HMM approach lies in its temporal modeling. Perhaps the major weakness of conventional HMMs is the modeling of state duration. In this report a connection between implicit state duration modeling in HMMs, explicit state duration modeling, and time invariant linear systems will be given. The work takes place in the context of the Ircam score follower while most approaches are given in speech recognition. The maximum state duration measured as the number of observations in speech recognition is typically 32 frames of 10 ms. The maximum state duration of a note is typically much higher. This has to be taken in account in the temporal modeling of the HMM of the score follower.