Describir: Visual music transcription of clarinet video recordings trained with audio-based labelled data