We will first look at the accuracy and loss during the training. Start TensorBoard, pointing to the log directory /tmp/tflearn_logs (default for tflearn): 

We find that both the validation and training loss decrease with the time steps. Note that the validation set we used was the same as from training, which was just a quick hack. You can set aside a separate validation set from the original data like we did for the test set. Next, we will look at the graph of the model in TensorBoard, as shown in the following screenshot:

As explained previously, we have an input layer that feeds the MFCC audio features tensor to the LSTM layer. The output of the LSTM is then fed to a fully connected layer that outputs the predictions. Next, we will look at creating a speech to text model using the DeepSpeech architecture.