# TODO

Planned work for this repository.

1. Train the adapter on the prepared LibriSpeech corpus.
2. Track reconstruction quality at each saved checkpoint via a held-out validation set, retaining only checkpoints that improve.
3. Report the final audio-reconstruction benchmark on a held-out test split (PESQ, STOI).
4. Publish the inverse-bijection decoder for recovering audio from generated spectrograms.
5. Release weights, decoder, and a representative set of demonstration audio samples accompanying the model card.