End-to-end automatic speech recognition system implemented in TensorFlow.

• Support TensorFlow r1.0 (2017-02-24)
• Support dropout for dynamic rnn (2017-03-11)
• Support running in shell file (2017-03-11)
• Support evaluation every several training epoches automatically (2017-03-11)
• Fix bugs for character-level automatic speech recognition (2017-03-14)
• Improve some function apis for reusable (2017-03-14)
• Add scaling for data preprocessing (2017-03-15)
• Add reusable support for LibriSpeech training (2017-03-15)
• Add simple n-gram model for random generation or statistical use (2017-03-23)
• Improve some code for pre-processing and training (2017-03-23)
• Replace TABs with blanks and add nist2wav converter script (2017-04-20)
• Add some data preparation code (2017-05-1)

If you want to replace feed dict operation with Tensorflow multi-thread and fifoqueue input pipeline, you can refer to my repo TensorFlow-Input-Pipeline for more example codes. My own practices prove that fifoqueue input pipeline would improve the training speed in some time.

If you want to look the history of speech recognition, I have collected the significant papers since 1981 in the ASR field. You can read awesome paper list in my repo awesome-speech-recognition-papers, all download links of papers are provided. I will update it every week to add new papers, including speech recognition, speech synthesis and language modelling. I hope that we won't miss any important papers in speech domain.

All my public repos will be updated in future, thanks for your stars!

• TensorFlow
• Numpy
• Scipy
• scikit-learn
• scikits.audiolab
• Kaldi recipe for wsj corpus (preprocessing stage)

This is a powerful library for automatic speech recognition, it is implemented in TensorFlow and support training with CPU/GPU. This library contains followings models you can choose to train your own model:

• Data Pre-processing
• Acoustic Modeling
  • RNN
  • BRNN
  • LSTM
  • BLSTM
  • GRU
  • BGRU
  • Dynamic RNN
  • Deep Residual Network
  • Seq2Seq with attention decoder
  • etc.
• CTC Decoding
• Evaluation(Mapping some similar phonemes)
• Saving or Restoring Model
• Mini-batch Training
• Training with GPU or CPU with TensorFlow
• Keeping logging of epoch time and error rate in disk

python train.py
	--mfcc_dir '/data/mfcc/'
	--label_dir '/data/label/'
	--keep False
	--save True
	--evaluation False
	--learning_rate 0.001
	--batch_size 32
	--num_feature 39
	--num_hidden 128
	--num_classes 28
	--save_dir '/src/save/'
	--restore_from '/src/save/'
	--model_checkpoint_path '/src/save/'

Instead of configuration in command line, you can also set the arguments above in train.py in practice.

The original TIMIT database contains 6300 utterances, but we find the 'SA' audio files occurs many times, it will lead bad bias for our speech recognition system. Therefore, we removed the all 'SA' files from the original dataset and attain the new TIMIT dataset, which contains only 5040 utterances including 3696 standard training set and 1344 test set.

Automatic Speech Recognition transcribes a raw audio file into character sequences; the preprocessing stage converts a raw audio file into feature vectors of several frames. We first split each audio file into 20ms Hamming windows with an overlap of 10ms, and then calculate the 12 mel frequency ceptral coefficients, appending an energy variable to each frame. This results in a vector of length 13. We then calculate the delta coefficients and delta-delta coefficients, attaining a total of 39 coefficients for each frame. In other words, each audio file is split into frames using the Hamming windows function, and each frame is extracted to a feature vector of length 39 (to attain a feature vector of different length, modify the settings in the file timit_preprocess.py.

In folder data/mfcc, each file is a feature matrix with size timeLength*39 of one audio file; in folder data/label, each file is a label vector according to the mfcc file.

If you want to set your own data preprocessing, you can edit calcmfcc.py or timit_preprocess.py.

Since the original TIMIT dataset contains 61 phonemes, we use 61 phonemes for training and evaluation, but when scoring, we mappd the 61 phonemes into 39 phonemes for better performance. We do this mapping according to the paper Speaker-independent phone recognition using hidden Markov models. The mapping details are as follows:

TODO

TODO

• dynamic RNN(GRU, LSTM)
• Residual Network(Deep CNN)
• CTC Decoding
• TIMIT Phoneme Edit Distance(PER)

• Add Attention Mechanism
• Add more efficient dynamic computation graph without padding
• List experimental results
• Implement more ASR models following newest investigations
• Provide fast TensorFlow Input Pipeline

If my code is helpful to you, please give me a star and fork to encourage me to keep updating. Thank you.

For any questions, welcome to send email to :[email protected]. If you use wechat, you can follow me by searching wechat public media id:deeplearningdigest, I would push several articles every week to share my deep learning practices with you. Thanks!