Melody transcription via generative pre-training