update readme for timit

README.md

@@ -12,7 +12,8 @@ They're not only for demonstration of the framework -- you can train them and re
 + [Generative Adversarial Network(GAN) variants (DCGAN,Image2Image,InfoGAN)](examples/GAN)
 + [Deep Q-Network(DQN) variants on Atari games](examples/Atari2600)
 + [Asynchronous Advantage Actor-Critic(A3C) with demos on OpenAI Gym](examples/OpenAIGym)
-+ [char-rnn language model](examples/char-rnn)
++ [LSTM-CTC for speech recognition](examples/TIMIT)
++ [char-RNN language model](examples/char-rnn)
 
 ## Features:
 

examples/README.md

@@ -22,4 +22,5 @@ Training examples with __reproducible__ and meaningful performance.
 + [Generative Adversarial Network(GAN) variants (DCGAN,Image2Image,InfoGAN)](examples/GAN)
 
 ## Speech / NLP:
++ [LSTM-CTC for speech recognition](TIMIT)
 + [char-rnn for fun](char-rnn)

examples/TIMIT/README.md

+
+## LSTM + CTC on TIMIT speech recognition dataset
+
+### Install Dependencies:
++ python binding for `lmdb`
+	+ `pip install --user lmdb`
++ `bob.ap` package for MFCC extraction
+	+ install blitz and openblas as dependencies of bob.ap
+	+ `pip install --user bob.ap`
+
+### Prepare Data:
+Convert NIST wav format to RIFF wav format:
+```
+cd /PATH/TO/TIMIT
+find -name '*.WAV' | parallel -P20 sox {} '{.}.wav'
+```
+
+Extract MFCC features and save everything to LMDB database:
+```
+./create-lmdb.py build --dataset /PATH/TO/TIMIT/TRAIN --db train.mdb
+./create-lmdb.py build --dataset /PATH/TO/TIMIT/TEST --db test.mdb
+```
+
+Compute mean/std of the training set (and save to `stats.data` by default):
+```
+./create-lmdb.py stat --db train.mdb
+```
+
+### Train:
+```
+./train-timit.py --train train.mdb --test test.mdb --stat stats.data
+```
+
+### Results:
+To be done

examples/TIMIT/train-timit.py

@@ -94,7 +94,7 @@ def get_config(ds_train, ds_test):
         ]),
         model=Model(),
         step_per_epoch=step_per_epoch,
-        max_epoch=500,
+        max_epoch=300,
     )
 
 if __name__ == '__main__':