CTC Loss
A specialized loss function used heavily in speech recognition to perfectly align an input audio sequence with an output text sequence of a different length.
Think of It Like This
Like stretching and squishing a rubber band with writing on it so the words perfectly line up with a completely different set of markings on the table.
In tasks like transcribing spoken audio to text, the model doesn't know exactly when a specific word was spoken in the audio file. Connectionist Temporal Classification solves this by introducing a 'blank' token and summing over all possible valid alignments. This allows the network to predict the final transcript without requiring expensive, frame-by-frame human alignment data.