Fast Signal Reconstruction from Magnitude Spectrogram of Continuous Wavelet Transform based on Spectrogram Consistency

Tomohiko Nakamura (The University of Tokyo) and Hirokazu Kameoka (The University of Tokyo/NTT Communication Science laboratories)

Demonstration of signal reconstruction from magnitude spectrogram

We here show audio examples of the proposed method, which can quickly reconstruct a time-domain signal from a magnitude spectrogram obtained with a continuous wavelet transform (CWT) [1]. These examples are from the RWC music database [2]. As the iteration number increases, the audio quality of the reconstructed signals improves more.
Original acoustic signal Magnitude CWT spectrogram with random phases Estimation results by the proposed algorithm
After 1 iterations
 After 10 iterations
 After 20 iterations
After 50 iterations
 After 100 iterations

Demonstration of pitch transposition

We here show audio examples of the proposed method. The pitch transposition is achieved by the proposed method, whereas only shifting the complex CWT spectrograms fail to the separation.
Original acoustic signal Shifting complex CWT spectrogram Shifting magnitude CWT spectrogram and zero phases Shifting magnitude CWT spectrogram and estimated phases
Convert the original into low pitch
Convert the original into high pitch

References

[1] Tomohiko Nakamura and Hirokazu Kameoka, “Fast signal reconstruction from magnitude spectrogram of continuous wavelet transform based on spectrogram consistency,” in Proceedings of International Conference on Digital Audio Effects, Sep. 2014, pp. 129–135.
paper , demo , [Travel Grant by the Hara Research Foundation]
[2] Masataka Goto, "Development of the RWC Music Database," Proc. 18th International Congress on Acoustics (ICA 2004), pp. I-553--556, April 2004.