Learning Transposition-Invariant Interval Features from Symbolic Music and Audio