Have you tried providing multiple pages at a time to the model? It might do better transcription as it have bigger context to work with.