Librispeech ASR model

The following models are provided: (i) TDNN-F based chain model based on the tdnn_1d_sp recipe, trained on 960h Librispeech data with 3x speed perturbation; (ii) Language models RNNLM trained on Librispeech trainiing transcriptions; and (iii) an i-vector extractor trained on a 200h subset of the data. We don't additionally include an LM since it can be prepared easily using the package available on OpenSLR. A tutorial on using the pre-trained model on your own data (with WSJ as an example) can be found here.