Turn Common Voice into a tf.data.Dataset
Currently we are using a stack of generators that chop up the sample from any folder containing files of any size. Besides the added complexity, we do not need this kind of flexibility and we lose a lot of meta data in CV.
Benefits:
- Using tf.data.Dataset we can leverage async. dataloading and more
- Using tf.data.Dataset we can simplify the training procedure using keras.fit()
- We will need meta data such as accent and age
- The dataset can be serialised and may be then used in a hackathon or else