Harmonic-Temporal Factor Decomposition (HTFD) Incorporating Music Prior Information for Informed Monaural Source Separation [1]


Authors: Tomohiko Nakamura(The University of Tokyo), Hirokazu Kameoka (Graduate School of Information Science and Technology, The University of Tokyo / NTT Communication Science Laboratories, Nippon Telegrah and Telephone Corporation).


Demonstration: Musical Key Conversion

We show a demonstration on automatic key transposition by using the proposed method method, or HTFD. The aim of key transposition is to change the key of a musical piece to another key. To do this, we

  1. separated the wavelet spectrogram of a music signal into the components associated with individual pitches using HTFD,
  2. transposed the pitches of a particular set of the separated components,
  3. added all the together to obtain the modified spectrogram,
  4. and constructed a time-domain signal from the modified spectrogram by using a fast signal construction algorithm from a wavelet power spectrogram , which we proposed previously [2].
For the comparision, we also carried out key transposition using harmonic NMF.


Original Audio signal Converted Audio signal obtained with HTFD Converted Audio signal obtained with harmonic NMF
RWC Classic Database No. 1, G major key
G natural minor key
G natural minor key
RWC Classic Database No. 36, E major key
E natural minor key
E natural minor key
The original audio files from RWC music database.

[1] Tomohiko Nakamura, Kotaro Shikata, Norihiro Takamune and Hirokazu Kameoka, "Harmonic-Temporal Factor Decomposition Incorporating Music Prior Information for Informed Monaural Souce Separation," Proc. the 15th International Society for Music Information Retrieval Conference (ISMIR2014), pp. 623--628, Oct. 2014.
[2] Tomohiko Nakamura and Hirokazu Kameoka, "Fast Signal Reconstruction from Magnitude Spectrogram of Continuous Wavelet Transform based on Spectrogram Consistency," Proc. 17th International Conference on Digital Audio Effects (DAFx-14), pp. 129--135, Sep. 2014.