Acoustic model training for non-audible murmur recognition using transformed normal speech data