Detection and transcription of drum hits
from audio files is a problem,
currently without an optimal solution.
Multiple methods give satisfactory results
but perfect transcription is hard to achieve
because of lack of information in the digital recording.
Approach using non-negative matrix factorization
assumes that we have access to isolated recordings
of every drum sound we wish to detect.
Short-term Fourier transform yields separate
transformations for each time frame. Isolated drum
spectrograms are used for non-negative matrix
factorization, the result of which we can then use for onset detection
and signal synthesis.
A transcription system for three different
drum sounds was implemented. We tested the system in
two separate scenarios and compared the results.
|
