HidekiKawahara
commented
1 month ago Dear Ikuma-san,
Thank you for your mail. In short, the GitHub implementation is correct. The article you cited is obsolete. The reference to the GitHub implementation corresponds to the following article, although it does not have detailed descriptions.
Kawahara, H., de Cheveigné, A., Banno, H., Takahashi, T., & Irino, T. (2005). Nearly defect-free F0 trajectory extraction for expressive speech modifications based on STRAIGHT. Proc. Interspeech, 537–540.
Thank you again.
Best regards, Hideki
2024/08/15 3:32、Kesh Ikuma @.***>のメール:
Kawahara-san, Thank you for making the STRAIGHT available on GitHub. Seeing its F0 detection performance in a 2022 paper by Vaysse et al. [2], I'm currently porting exstraightsource() function to Python so I can run more in-depth analysis with pathological voice data. In doing so, I found a likely discrepancy in the generation of the filter impulse response in multanalytFineCSPB() Lines 41-49. In Eurospeech 1999 paper [1], you have a trio of equations which correspond to these lines: $$ \begin{align} w_s(t,\lambda) &= w(t,\lambda) ⋆ h(t,\lambda) \ w(t,\lambda) &= \exp\left(-\frac{\lambda^2t^2}{4\pi \eta^2}\right) +\exp(j\lambda t) \ h(t,\lambda) &= \begin{cases} 1-\left| \frac{\lambda t}{2\pi\eta} \right| & \text{if } \left| t \right| > \frac{2\pi\eta}{\lambda}\ 0 & \text{otherwise} \end{cases} \end{align} $$ However, the MATLAB code appears to implement: $$ w_s(t,\lambda) = \left[ \exp\left(-\frac{\lambda^2 \tau^2}{4\pi \eta^2}\right) ⋆ h(\tau,\lambda) \right] \exp(j\lambda t) $$ Is this an evolution of the algorithm from '99 to '18 or a bug? Thank you Kesh Ikuma [1] H. Kawahara, H. Katayose, A. D. Cheveigné, and R. D. Patterson, “Fixed point analysis of frequency to instantaneous frequency mapping for accurate estimation of F0 and periodicity,” in 6th European Conference on Speech Communication and Technology (Eurospeech 1999), ISCA, Sep. 1999, pp. 2781–2784. doi: 10.21437/Eurospeech.1999-613. [2] R. Vaysse, C. Astésano, and J. Farinas, “Performance analysis of various fundamental frequency estimation algorithms in the context of pathological speech,” Journal of the Acoustical Society of America, vol. 152, no. 5, pp. 3091–3101, 2022, doi: 10.1121/10.0015143. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>
Hideki Kawahara, Emeritus Professor, Wakayama University, Japan https://web.wakayama-u.ac.jp/~kawahara/index_e.html
tikuma-lsuhsc
Kawahara-san,
Thank you for making the STRAIGHT available on GitHub. Seeing its F0 detection performance in a 2022 paper by Vaysse et al. [2], I'm currently porting
exstraightsource()function to Python so I can run more in-depth analysis with pathological voice data. In doing so, I found a likely discrepancy in the generation of the filter impulse response inmultanalytFineCSPB()Lines 41-49. In Eurospeech 1999 paper [1], you have a trio of equations which correspond to these lines:$$ \begin{align} w_s(t,\lambda) &= w(t,\lambda) ⋆ h(t,\lambda) \ w(t,\lambda) &= \exp\left(-\frac{\lambda^2t^2}{4\pi \eta^2}\right) +\exp(j\lambda t) \ h(t,\lambda) &= \begin{cases} 1-\left| \frac{\lambda t}{2\pi\eta} \right| & \text{if } \left| t \right| > \frac{2\pi\eta}{\lambda}\ 0 & \text{otherwise} \end{cases} \end{align} $$
However, the MATLAB code appears to implement:
$$ w_s(t,\lambda) = \left[ \exp\left(-\frac{\lambda^2 t^2}{4\pi \eta^2}\right) ⋆ h(t,\lambda) \right] \exp(j m \lambda t) $$
where $m$ is the harmonic number under assessment.
Is this an evolution of the algorithm from '99 to '18 or a bug or am I overlooking something?
Thank you Kesh Ikuma
[1] H. Kawahara, H. Katayose, A. D. Cheveigné, and R. D. Patterson, “Fixed point analysis of frequency to instantaneous frequency mapping for accurate estimation of F0 and periodicity,” in 6th European Conference on Speech Communication and Technology (Eurospeech 1999), ISCA, Sep. 1999, pp. 2781–2784. doi: 10.21437/Eurospeech.1999-613. [2] R. Vaysse, C. Astésano, and J. Farinas, “Performance analysis of various fundamental frequency estimation algorithms in the context of pathological speech,” Journal of the Acoustical Society of America, vol. 152, no. 5, pp. 3091–3101, 2022, doi: 10.1121/10.0015143.