CTC-TIMIT

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.extension bob.blitz bob.core bob.sp bob.ap

Prepare Data:

We assume the following file structure:

TRAIN/
  DR1/
    FCJF0/
      *.WAV     # NIST WAV file
      *.TXT
      *.PHN
  ...

Convert NIST wav format to RIFF wav format:

cd /PATH/TO/TIMIT
find -name '*.WAV' | parallel -P20 sox {} '{.}.wav'

Extract MFCC features and phoneme labels, and save everything to LMDB database. The preprocessing follows the setup in

• Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with RNN - Alex Graves

./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:

Get 0.28 LER (normalized edit distance) after about 40 epochs.