Fig. 2

Stages for training a hybrid DNN-HMM acoustic model on the top of a triphone-based, speaker-adapted GMM one. Lexicon and LM are omitted, but are basic inputs for all training blocks as well as the MFCCs. Scripts are based on Kaldi’s Mini-librispeech recipe, and the DNN in particular was trained following the nnet3 setup. At the top of each block, there is a reference for a script (boldface) and a directory where resources created by that script are usually stored (normal text). More details on the DNN training (dashed block) are depicted in Fig. 3