Create README.md

Tests/EndToEndTests/Speech/LSTM_CTC/README.md

@@ -0,0 +1,18 @@
+CNTK implementation of Connectionist Temporal Classification (CTC) is based on the paper by Alex Graves, etl. "Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks".  
+These instructions are written in the context of speech recognition, yet remain relevant to handwriting recognition and other domains.
+The criterion node for CTC training is *ForwardBackward*. This node is an abstraction for different types of training methods based on the forward/backward Viterbi-like pass. The node takes as the input the graph of labels, produced by the *LabelsToGraph* node that determines the exact forward/backward procedure. The *EditDistanceError* node is used for error evaluation. Here is [example CTC training configuration](https://github.com/Microsoft/CNTK/blob/master/Tests/EndToEndTests/Speech/LSTM_CTC/lstm.bs).
+
+# Preparation of Training Data
+
+Features are consumed with the HTKMLFDeserializer and labels with the TextDeserializer. To convert a traditional set of MLF/SCP files to the format suitable for CTC training, use [this python script](https://github.com/vmazalov/CNTKConfigs/blob/master/ctc_label_conversion.py). Assuming that the python script is placed into the CNTK directory with speech test data, example run to convert features is
+<pre><code> %CNTK_path%\Tests\EndToEndTests\Speech\Data>python ctc_label_conversion.py --inputScpFile glob_0000.scp --inputMlfFile glob_0000.mlf --inputPhoneListFile state.list --outputScpFile ctc_glob_0000.scp --outputLabelFile ctc_glob_0000.mlf </code></pre>
+
+There are several assumptions about the input data:
+
+* The CTC blank token is expected to be the last label in the label sequence. For the referenced example, the blank token is a label with index 132.
+
+*  The id of the blank token should be provided in the ForwardBackward node and as the tokenToIgnore in the EditDistanceError node.
+
+* The label file (consumed by TextDeserializer) is expected to have the following format: 
+  *seqId |l phoneId:[1|2]*,
+where seqId is the id of the sequence/utterance to which given frame belongs, phoneId is index of the phone of given frame, phoneId is followed by ":1" if this is not the first frame of the phone, and phoneId is followed by ":2" if this is the first frame of the phone.