6.5.2 (Page 2 of 2)

Hi Gilles,
Thank you for this review.
The correction made on the lower notes in your fxp with your tool brings a deeper sound also very nice.

By analyzing the spectrum profile of your fxp on these notes and comparing it to mine, I noticed several differences:
(1) Much more contrast, explainable difference because I used here a scale factor dividing (in db) by 4 the spectrum differences initially obtained. Your fxp shows me that the scale factor I use must be less important at least for the lowest notes. (With a scale factor that divides by only 1.1 on these low notes, I can get a sound closer to the original one (scale factor, however, unsuitable for higher notes). Unfortunately, I have not yet enough perspective to choose the best scale factor note by note automatically according to the frequency of the fundamental.In my fxp, the scale factor was constant: division (in db) by 4 no manually modified, which should be reviewed)
(2) Differences sometimes completely opposite on the level to be added or subtracted on this or that partial.
Is this due to differences in methods used for spectral analysis between your tool and my pieces of code? (at your disposal if you wish).
In practice I use (via Audacity) a spectral analysis with a Blackman-Harris type window, whose size varies according to the range of notes according to the rule (empirical): next, maybe to re-adapt: A- 1 to F1 => 65k with the 686 ms from the beginning of the sample;
F # 1 to F2 => 32k with the 346 ms of the beginning .., F # 2 to D # 4 => 16k with the 172ms of the beginning; E4 to F5 => 8k with the 88ms of the beginning; F # 5 to B5 => 4k with the beginning 46ms; C6 to B6 => 2k with the 24ms of the beginning; C7 => 1k with the 14ms of the beginning.
The results between B0 and B3 are of insufficient quality with this empirical technique (nevertheless preserved in the state in the diffused fxp).
I think you use longer samples (1s) and a 16 or 32k rectangular window with your tool, is that it?
For my part, I had avoided the use of rectangular windows because I found (with Audacity) a lower discriminating power above 4khz (not because of the frequency itself, but because the level is much lower for high harmonics), but I'm not at all certain to have chosen the right type of window, the most suitable, nor the optimal sample size, what do you think?
I had only noticed, the higher harmonics disappearing quickly, it was not necessary that the duration of the sample is not too important after the impact of the hammer ..., unfortunately in this case to the detriment of the level of the first harmonics for the bass.

Would you be able to reproduce with your tool in a fxp the following notes of my recording of T130 with 2 microphones SE8, I hope that it will be able to make disappear the horrible stamp which I obtain for example on C # 2 and G # 2 for example ...
NB: for the reconciliation of real and virtual notes, my starting point was also a fxp with only 2 microphones.

Bruno

I use the scale factor very empirically in the note edit window: I simply adjust it by ear to get rid of unwanted artefacts, usually overemphasized bell sound.
In this case A-1 was left as is but on E0 and F0 I had to rescale a lot partly because the source and destination notes were not exactly in tune (to my ears)

Edit: Something just hit me that explains the spectrum differences: I started from your already modified spectrum to generate mine, while you started from the raw U4, so what I get is a sort of combination of your changes with mine...The result is then not exact since pasting the note edit string replaces your result and does not add to it. I guess I should start again by removing your spectrum values first...sigh...

I am by no mean an expert in signal processing...I simply used an FFT algorithm written in Java I found on the web and built a specialized program around it for use with Pianoteq Pro. I experimented with it in the past and found that the default 16K sample number and rectangular window seemed to work best for the type of signal piano notes provide, mostly short, so the window limits effects are not that important in my opinion. I used for the Schimmel your longer samples (4.5s) instead of the short ones (0.4s) because they were easier to extract...(lazy...) but trying for instance A-1 with your short sample and a 65K window only produced more bell inharmonics. This is all very empirical because for instance the signal starts with a hammer sound that is separate in Pianoteq and thus wrongly included in the sound, and also there is, when recording long notes, the sound of the damper also modeled separately. I think the only justification for my approach (or yours) is that we get audibly closer to the target sound. We can never reach it though... My software only accepts 16/44.1 samples so I also am less concerned by FFT window size or type than you if you work with bigger 24/96 samples.

I don't hear anything wrong with G#2 but the "zing" resonance on C#2 can be removed by changing the string length to something like 1.03m. Using my tool only produce a very dull result. I had to cut the note short because of the important damper sound at the end.