Non-negative matrix factorization based compensation of music for automatic speech recognition