Tensorflow implementation of Phoneme Recognition using Connectionist Temporal Classification. http://www.cs.toronto.edu/~graves/icml_2006.pdf. The model is created by bidirectional dynamic rnn with three stacked lstms in each direction.
First, transform the Nist wav file in TIMIT dataset to normal wav file. I use the scikits.audiolab to finish this task.
python phoneme_ctc.py transform -n train_dir -t test_dir -o out_dir
I extract the mfcc features using the module python_speech_features.
The dimension of feature is 39 and there are 39 classes for output phoneme labels.
The work flow is:
extracted features from wav audio -> network -> phoneme labels.
After 100 epoches, which costs me 4 days using NVIDIA1070 GPU, the cost decrease from 120 to 40 and the label error rate from 90% to 30%.
python phoneme_ctc.py train -n processed_train_dir -t processed_test_dir -o checkpoint_dir
python phoneme_ctc.py decode -m checkpoint_dir. Then you can provide a wav audio file and the model will output the phoneme labels. Notce: the wav file should be 1 channel, 16 bits and bitrate 16000. I record the wav file using sox on my mac. e.g. sox -d -b 16 -c 1 -r 16000 helloworld.wav
I provide a trained model, you can test the model using python phoneme_ctc.py decode -m model.
python 2.7
brew install libsndfile
pip install -r requirements.txt