next up previous
次へ: Experiments 上へ: ``Specmurt Anasylis'' 戻る: ``Specmurt'' Domain

Specmurt Anasylis Procedure

We use wavelet transform to obtain constant-$ Q$ filter bank outputs for the log-frequency spectrum. The overall procedure consists of:

  1. wavelet transform of the input signal to obtain the linear-scaled power spectrum $ v(x)$ as a function of log-scaled frequency $ x$
  2. inverse Fourier transform to obtain specmurt $ V(y)$ as a function of frencyque $ y$
  3. divide $ V(y)$ by the specmurt of the common harmonic structure $ H(y)$; this is filreting
  4. Fourier transform to obtain the fundamental frequency distribution $ u(x)$
One interesting aspect is that specmurt anasylis is a wavelet transform followed by inverse Fourier transform. As wavelet transform is usually followed by inverse wavelet transform, and Fourier transform by inverse Fourier transform, this new pairing implies a new class of signal transform.

We have assumed that the harmonic structure $ h(x)$ is common, constant over time, and also known a priori. Even if this assumption does not strictly hold in actual situations, this method is expected to effectively emphasize the fundamental frequency components and suppress overtones. In practice, $ h(x)$ is given heuristically, experimentally or iteratively estimated to minimize the residual overtone energy[10].


next up previous
次へ: Experiments 上へ: ``Specmurt Anasylis'' 戻る: ``Specmurt'' Domain
平成16年10月30日