PietW. wrote:It is important the sample to tune to 172 Hz. Otherwise, the resynthesis will not function correctly.
Indeed the quality depends on correct tuning. The sample analysis works as follows: AT does an FFT with 1024 samples (= 512 partials). Then it devides the result into 33 ranges (= the slots or slices).
If the source sample is too short for filling a complete table, it stretches the analysis result. If it's too long, a number of the source data is summarized to a certain slot. This way you don't have to have a certain sample length to fill a complete table. In the latter case the SF_MODE button appears. This button disables the summarizing, which in some cases can produce better results. By the way, the "optimum" would be a sample of 172.265626 Hz fundamental frequency and a length of 33792 samples.
Because the Komplexer engine only uses 64 partials, these 64 partials have to be extracted from the 1024 point FFT. The "tuning keyboard" on the "SAMPLE ANALYSIS OPTIONS II" page sets the root frequency in a range of +/- 1 octave around the ideal value of 172.265625 Hz and the corresponding overtones are weighted and extracted from the raw analysis data. This is quite important on harmonic sounds - on speech or similar "noisy" material it's less important. The arbitrary phases of the source material are recalculated to +/-.
Using a smaller FFT to get the 64 partials directly has two disadvantages:
1.) The FFT window will become more audible, because the resynthesis cannot use overlap/add to cancel out artifacts.
2.) It would be necessary to tune every sample exactly before analysis (which would take more time to find the best settings, since you'd have to do the resynthesis again and again).
The results will often sound very "mechanic" because of the reduction to the 64 partials and the limitation to sine waves. This is far from what most source material consists of. This is also the reason for the different look (and in some cases different sound) of imported single cycles: Everything is analyzed and rebuild from sines only.
One advantage of removing arbitrary phases in wavetable synthesis is that phase shiftings between diffrent slots don't "take over" the pitch: For a short single cycle the audible pitch is defined by it's loop length. But regarding longer samples pitch is also defined as a partial's phase increment over time. This can lead to a detuned sounding wavetable when using a fast lfo or envelope on wavetables which holds sine and
consine content (although allowing arbitrary phases would allow a wider range of sounds).
Regarding the need for nomalizing the amplitudes of the partials, especially after using the RANDOM feature: As explained some posts above the audio output has to be limited to a certain dynamic range. This means that the amplitudes of a slice have to be recalculated to "what you hear" - in most cases they have to be lowered.
Without limiting a saw wave's amplitude would be up to 50 times the amplitude of a single sine. Unfortunately the proportion of partial and wave amplitudes depends on polarity and partial frequency, this is what makes handling amplitudes so difficult. Futher analyzed samples have to be treated differently from a simple list of single cycles and it took me months to get at least this - sometimes confusing - solution.