tensorflow-ctc-speech-recognition | Connectionist Temporal Classification | Speech library

Look closely at your input texts (rand_target), I'm sure you see some simple pattern which correlates with the inf loss value ;-)

A short explanation of what is happening: CTC encodes text by allowing each character to be repeated and it also allows a non-character marker (called "CTC blank label") to be inserted between characters. Undoing this encoding (or decoding) then simply means throwing away repeated characters and then throwing away all blanks. To give some examples ("..." corresponds to text, '...' to encodings and '-' to the blank label):

• "to" -> 'tttooo', or 't-o' or 't-oo', or 'to', and so on ...
• "too" -> 'to-o', or 'tttoo---oo', or '---t-o-o--', but NOT 'too' (think about how the decoded 'too' would look like)

Now we know enough to see why some of your samples fail:

• the length of your input text is 2
• the length of the encodings is 2
• if the input character is repeated (e.g. '11', or as a python list: [1, 1]), then the only way to encode this would be by placing a blank in between (think abound decoding '11' and '1-1'). But then the encoding would have a length of 3.
• so, there is no way to encode texts of length 2 with a repeated character into a length 2 encoding, therefore the TF loss implementation returns inf

You can also imagine the encoding as a state machine - see illustration below. The text "11" can be represented by all possible paths starting at a start state (two leftmost states) and ending at a final state (two rightmost states). As you can see, the shortest possible path is '1-1'.

To conclude, you have to account for at least one additional blank to be inserted for each repeated character in the input text. Maybe this article helps in understanding CTC: https://towardsdatascience.com/3797e43a86c