Phoneme Recognition using RecNet

This repository contains a automated phoneme recognition based on a recurrent neural network. The implementation uses the RecNet framework which is based on Theano. The used speech data set is the TIMIT Acoustic-Phonetic Continuous Speech Corpus.

Requirements

• The pre-process requires the SND file format library. It's installable via the APT interface.

sudo apt-get install libsndfile-dev

• There's a problem in 'scikits.audiolab's setup.py file. Workaround: first update pip, second install numpy, then requirements
• Furthermore the model requires the packages listed in the requirements.txt.

pip install -r requirements.txt

• This phoneme recognition uses the RecNet framework which needs to be installed.

git clone https://github.com/joergfranke/recnet.git
cd recnet
python setup.py install

Please find a proposal for setup phoneme recognition in setup.sh

How to use it

Step 1: Make a feature/target data set

This step contains the whole pre-process and creates a data set in the form of two lists, one with sequences of features (MFCC + log-energy + derivations) and one with corresponding sequences of targets (correct phonemes). This pre-process is orientated on Graves and Schmidhuber, 2005 . The data set gets stored in the klepto file format. Do the following for creating the data set:

1. Add path to the TIMIT corpus
2. Run make_data_set.py

Step 2: Train recurrent neural network model

The second step contains the training of the model. This phoneme recognition uses for instance gated recurrent units (GRU) with layer normalization. Do the following for training the model:

1. Run train_model.py
2. Find the log file of training in the log folder.

Step 3: Evaluate exercised model

At least the exercised model gets evaluated on the test set. The log loss and the rate of correct detected phonemes will calculated. A plot shows the input features, the true and the predicted phonemes.

1. Run evaluate_model.py