– Guest lecturer introduction
– Outline
– Modern speech recognition
– Connectionist temporal classification
@Alfredo Canziani if we know the log.probabilities (at time from the speech video), and the actual-softmax gives an approximation to the maximum log.probability, in this case why do we need to apply actual-softmax if we can just find the maximum log.probability without it? I understand what is actual-softmax, but in this case actual-softmax is only used because it is differentiable, and therefore it can be used to do backprop? Thanks
– Decoding with beam search (inference)
