notes

README.md

@@ -1,14 +1,16 @@
 # stanford-ctc
 Neural net code for lexicon-free speech recognition with connectionist temporal classification
 
-This repository contains code for a bi-directional RNN training using the CTC loss function. 
+This repository contains code for a bi-directional RNN training using the CTC loss function.
 We assume you have separately prepared a dataset of speech utterances with audio features and text transcriptions.
 
 For more information please see the [project page](http://deeplearning.stanford.edu/lexfree/) and the [character language modeling repository](https://github.com/zxie/nn)
 
 Our neural net code runs on the GPU using [Cudamat](https://github.com/cudamat/cudamat)
 We use a forked version of Cudamat to add an extra function which you can find [here](https://github.com/awni/cudamat). If you need a more recent version of cudamat you can likely take just the extra function and apply the patch to the most recent version of Cudamat.
 
+The latest code is in the directory `ctc_fast`; please set your `PYTHONPATH` accordingly. The script `runNNet.py` should be the starting point for training the BRNN model -- you'll have to modify `run_cfg.py` and `decoder_config.py`. Unfortunately the `run*.sh` scripts in `{timit/wsj/swbd}-utils` are outdated but you can refer to them for reasonable parameter settings.
+
 Example feat#.bin, keys#.txt, and alis#.txt files for small subset of TIMIT training data can be
 found [here](http://deeplearning.stanford.edu/lexfree/timit/).