Tensorflow implementation of models trained for language classification on Mozilla Common Voice using tf.Dataset API.
The offline_process.py script converts the .mp3 files of dataset to a .tfrecord file per language. Each of the .tfrecord files contain an array of tuples of an MFCC spectrogram for audio clip and the corresponding label string.
This method saves a lot of computation since the mp3 decoding and pre-processing is done only once.
The dataset.py implements the dataset pipeline where mp3 files are decoded and processed to features on demand. This is very computationally expensive. If you plan to do multiple training runs or hyper-parameter tuning, please use the TFRecord method above.
Tested on tensorflow==2.1.0. To train an LSTM model run:
python3 train.py
This should achieve ~85% accuracy in 10 epochs.