I'm trying to come up to speed on the use of deep neural networks for speech recognition and I'm confused about what exactly are the targets for the DNN. I understand that the DNN tries to predict a series on senones, ie the inputs to the network are a window of audio signal features (like MFCCs and their derivatives) and the output is a one-hot vector indicating the senone. But to train this, we must have a set of targets in the training data, ie targets that we get from somewhere else. From what I understand (I could be wrong of course), these targets generally are obtained from using some other existing system like an HMM-GMM based speech recognition system. But this doesn't make sense to me - that HMM-GMM system will have its own errors, so wouldn't using those senones as ground truth make the DNN based system only as good as the HMM-GMM system? How can a DNN improve on the HMM-GMM system in this case? Can someone please explain this?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.