The task is to do a speech-to-text. Check out this tutorial: https://www.kaggle.com/kingabzpro/wave2vec-asr-wolof

Yes as already noted, the challenge is about speech to text. Until recently, people would most likely have used a seq2seq model (for example a bi-directional LSTM). However, around 2 years ago Facebook did quite some work in this domein resulting in their wav2vec2.0 model which is today state of the art.

For your questions if all characters are represented in the training set: not necessarily. Usually you assume the same distribution of the training and test set but this does not striclty imply that you would observer the whole vocabulary.

For using external data: you would need to check the competition rules. However, the models I have mentioned above are usually not trained from scratch as the resource requirements are too high.