An example trying to run baidu asr system "https://github.com/baidu-research/ba-dls-deepspeech" using the default ctc function tensorflow. Similar to the Baidu system I am using LibriSpeech dataset. The project is not complete.

As of Feb 22, 2017

• The GPU I am using are running out of memory. ( tensorflow/core/common_runtime/gpu/pool_allocator.cc:247] PoolAllocator) Pool allocator is taking a lot of time if I train it on more than few 100's of audio files.
• I am still testing, so the Neural Network is not deep. I am using only train dataset and no test or validation data sets.

I made use of the LibriSpeech ASR corpus. How to download dataset and create json files is described in the above mentioned baidu github. I used the audio files of speaker ID 19 to train my system.

ctc_tensorflow_example.py, ctc_tensorflow_multidata_example.py, utils.py are from the forked directory ctc_spectrogram_single.py downlaods one timit dataset wavefile and uses its spectrogram to train NN ctc_spectrogram_multidata.py takes multiple audio files as input (train_corpus.json) train_corpus.json > from baidu asr

• Python 2.7+
• Tensorflow 0.12
• python_speech_features (not necessary if using spectrogram)
• numpy
• scipy

This project is licensed under the terms of the MIT license.