For sure, there is a (well known) theoretical background for pitch estimation, including many interesting academic papers with comparative studies of methods. On the other hand, one knows that reverberant room effects can be handled through signal pre-whitening methods. Nonetheless, my question is to those who, like myself, feel frustrated by the almost erratic performance of pitch estimators in naturally spoken sentences (i.e. normal rhythm) in small reverberant rooms, even after signal pre-whitening. Thus, I would like to know is someone successfully experimented new pragmatic approaches, possibly not convencional ones.
Dear Yi Xu, thank you for your interest in my question, as well as for your useful answer. Indeed, ProsodyPro seems to be a powerful platform for prosody analysis. However I tried it with a specific recorded speech sample under natural reverberation and noise - the very same recording that first motivated my question - and I got the same kind of problem that I noticed in all methods I implemented and tested before. Namely, although the recording has a low-pitch male voice (around 100 Hz, as far as I can subjectively evaluate with my ears), all methods I've tried until now erroneously detect a persistent pitch above 200 Hz, with a temporal behaviour which is not that of a typical 'doubling' effect. I'm putting this problematic speech file in my dropbox, just in case you feel like trying it too.
https://dl.dropboxusercontent.com/u/1791564/problematicPitch.wav
The problem with the recording sample is not just room reverberation, which is not that severe, but mainly low signal-to-noise ratio. The likely reason is that the microphone was too far away from the speaker. I would say that it is hard for any pitch detection algorithm to do a reasonably good job on signals of this quality. But if pitch analysis is absolutely necessary in this case, ProsodyPro gives you a chance to do the best you can humanly do to get some reasonable reading. Please see the Excel plot in the attached zip file.
Dear Yi Xu,I appreciate a lot your taking some time to analyse my sound sample. Now, let me explain why I made such a supposition that room reverberation was the main problem I'm fighting against. First, I implemented and used a very basic autocorrelation based method, whose result is shown in the upper figure (appended pdf file). Although I can feel (with my ears) a lot of F0 signal (almost all the time) around 100 Hz, what the computer shows me is a lot of estimated F0 around 200 Hz. For sure, it could be the well-known doubling effect (first hypothesis). But the weird thing is that I also can feel (with my ears) like a room reverberation mode being excited around 200 Hz (which may be compatible with some small room's dimension), what leads me to the second hypothesis.
In case it is just a room reverberation effect, I do expect to reduce it through spectral subtraction (for it is expected to vary much less than the excitation spectrum). Therefore, I just subtracted the average correlation profile (thanks to the linearity property of the inverse Fourier transform of the average spectrum) and re-estimated the F0 with the subtracted window correlation profiles. The result is shown in the second figure, were a substantial improvement can be noticed, thus corroborating the room reverberation hypothesis.
Unfortunately, it depends on a post-processing procedure (for average profile estimation) and it, in spite of the noticed improvement, it does not yet yield all the F0 contour that I can detect with my ear (which is my ultimate goal).
Obs.: Please notice that the strong room effect is also in agreement with your remark: 'The likely reason is that the microphone was too far away from the speaker.'.
- Badges
- Science topic