About the formula for computing the observation probability in function log_densities in RNNBeatTrackingObservationModel:

[moonryul]

I have some issues about
madmom -> features -> beat_hmm.py => RNNBeatTrackingObservationModel => log_densities (marked by *** ...**** below):

def log_densities(self, observations):
"""
Compute the log densities of the observations.

    Parameters
    ----------
    observations : numpy array, shape (N, )
        Observations (i.e. 1D beat activations of the RNN).

    Returns
    -------
    numpy array, shape (N, 2)
        Log densities of the observations, the columns represent the
        observation log probability densities for no-beats and beats.

    """
    # init densities
    log_densities = np.empty((len(observations), 2), dtype=np.float)
    # Note: it's faster to call np.log 2 times instead of once on the     whole 2d array

  ****  
   log_densities[:, 0] = np.log((1. - observations) /
                                 (self.observation_lambda - 1))

    log_densities[:, 1] = np.log(observations) ****

    # return the densities
    return log_densities

=> The formula in *** ...****
(1. - observations) /
(self.observation_lambda - 1)
corresponds to eq (3)

P(a_k | phi_k) = a_k if 1<= phi_k <= PHI/lambda;
(1-a_k)/(lambda-1) otherwise

in paper "A MULTI-MODEL APPROACH TO BEAT TRACKING CONSIDERING
HETEROGENEOUS MUSIC STYLES", 2014.

But this observation prob is not normalized either in paper or in the code.
I could not find any place where this prob is normalized.

Could you tell me where this prob is normalized?
Or, is it OK that this observation prob is not normalized, say in the Viterbi algorithm?