githubharald

"""
Compute score for decoded text in a CTC-trained neural network using TensorFlow:
1. decode text with best path decoding (or some other decoder)
2. feed decoded text into loss function
3. loss is negative logarithm of probability

Example data: two time-steps, 2 labels (0, 1) and the blank label (2). 
Decoding results in [0] (i.e. string containing one entry for label 0).
The probability is the sum over all paths yielding [0], these are: [0, 0], [0, 2], [2, 0]
with probability