I tend to agree. Cutting out the extra processing and telling the sampler "actually, this is meant to sound this loud at this velocity, don't attenuate it" seems the simplest approach to me.Sascha Franck wrote:After all, I actually seem to prefer lower leveled samples for lower velocity zones - that way I don't have to fool with individual zone level adjustments myself.
Free multisampled conga in sfz format
-
- KVRAF
- 7412 posts since 8 Feb, 2003 from London, UK
-
- KVRian
- 500 posts since 13 Oct, 2004 from Durham, NC USA
I disagree, Sasha. It's not just impossible, it's what I did, and what I believe Scarbee did. Though not exactly the way you said.
What works better is to set the recording level anew for each velocity layer, and sample all the notes on that velocity layer using the same setting.
The default soundfont velocity scaling assumes that all samples are more or less normalized. (Some caveats about that, but let's not get too ridiculously technical here!) That's the reason for pljones's first post above.
The sample player will be adjusting the volume based on the velocity anyway. If all layers are normalized, then only attenuation is ever required. Furthermore, it makes it a lot easier to tune the dynamics of the soundfont using small adjustments -- or better yet, just finding the sweet spot in the MIDI velocity for each layer.
My quietest samples were about 25 dB down from the loudest. If I recorded them all at the same level, those quietest samples would effectively be 12-bit samples. Well, if I'm going to play a very quiet part for a whole song, I'd rather be using 16-bit samples than 12-bit samples! And note that it's only 12 bits for the attack -- the decay part will be far fewer bits. I know very well the sound of losing bit depth on decay, and I don't like it!
Note that a VSTi or DXi sample player will be running in 32-bit float mode, so you will get the extra bits of accuracy without doing anything special (with normalized samples). This wouldn't be true for a player built on a fixed-point DSP chip.
What works better is to set the recording level anew for each velocity layer, and sample all the notes on that velocity layer using the same setting.
The default soundfont velocity scaling assumes that all samples are more or less normalized. (Some caveats about that, but let's not get too ridiculously technical here!) That's the reason for pljones's first post above.
The sample player will be adjusting the volume based on the velocity anyway. If all layers are normalized, then only attenuation is ever required. Furthermore, it makes it a lot easier to tune the dynamics of the soundfont using small adjustments -- or better yet, just finding the sweet spot in the MIDI velocity for each layer.
My quietest samples were about 25 dB down from the loudest. If I recorded them all at the same level, those quietest samples would effectively be 12-bit samples. Well, if I'm going to play a very quiet part for a whole song, I'd rather be using 16-bit samples than 12-bit samples! And note that it's only 12 bits for the attack -- the decay part will be far fewer bits. I know very well the sound of losing bit depth on decay, and I don't like it!
Note that a VSTi or DXi sample player will be running in 32-bit float mode, so you will get the extra bits of accuracy without doing anything special (with normalized samples). This wouldn't be true for a player built on a fixed-point DSP chip.
-
- KVRAF
- 4143 posts since 7 Sep, 2001 from Melbourne, Australia
When you say all layers are normalised I'm getting a little confused.
I would think if you have velocity-switched samples you would want each velocity layer to peak at a certain decibel range. For the loudest layer that may be 0dB and all zones in that layer would be normalised to be consistent.
But then on other velocity layers wouldn't you set an arbitrary level of dB that all zones in that layer should peak at and then set them all to that level?
I don't think I understand the discussion about different bit levels either. 12 bits for the attack? What do you mean?
Caleb
I would think if you have velocity-switched samples you would want each velocity layer to peak at a certain decibel range. For the loudest layer that may be 0dB and all zones in that layer would be normalised to be consistent.
But then on other velocity layers wouldn't you set an arbitrary level of dB that all zones in that layer should peak at and then set them all to that level?
I don't think I understand the discussion about different bit levels either. 12 bits for the attack? What do you mean?
Caleb
Happiness is the hidden behind the obvious.
-
- KVRian
- 500 posts since 13 Oct, 2004 from Durham, NC USA
Brad:
Great first effort!
A few suggestions:
1) Trim the starts a lot closer. Your notes start as late as 25 msec into the recording. This causes two problems, latency and retriggering. You have the notes properly set up in a group so that playing one stops all the others. But notice the gap if you play the tone3 sample repeatedly. The 9 msec gap is audible and unnatural sounding.
2) Record the quieter velocities louder, as I mentioned above. It's a pretty quiet soundfont unless I'm playing all notes in the high velocity range, and mostly using Tone. (I was playing the sf2 version.)
3) Record both right and left hand samples. This may seem redundant, but it's not. I find that if I have a right and left hand sample and alternate them when playing a part with a repeated note (e.g., like a roll), it sounds far more natural than repeating the same note. Too bad whoever designed the GM drum map didn't know enough about programming drum parts to realize this and assign more keys to snare and hihat! Of course, you'd map the left hand notes to the lower octave. It might only be needed for the Tone samples, though -- it depends on how much conga players use the same kind of strike twice or more in a row.
4) RECORD THE OTHER CONGA! (
)
Regardless, your soundfont is way better than the conga built into my Ensoniq MR76 synth, which in general has great drum sounds. I'm sure folks will find excellent use for it.
Thanks again for sharing it!
Jeff
Great first effort!
A few suggestions:
1) Trim the starts a lot closer. Your notes start as late as 25 msec into the recording. This causes two problems, latency and retriggering. You have the notes properly set up in a group so that playing one stops all the others. But notice the gap if you play the tone3 sample repeatedly. The 9 msec gap is audible and unnatural sounding.
2) Record the quieter velocities louder, as I mentioned above. It's a pretty quiet soundfont unless I'm playing all notes in the high velocity range, and mostly using Tone. (I was playing the sf2 version.)
3) Record both right and left hand samples. This may seem redundant, but it's not. I find that if I have a right and left hand sample and alternate them when playing a part with a repeated note (e.g., like a roll), it sounds far more natural than repeating the same note. Too bad whoever designed the GM drum map didn't know enough about programming drum parts to realize this and assign more keys to snare and hihat! Of course, you'd map the left hand notes to the lower octave. It might only be needed for the Tone samples, though -- it depends on how much conga players use the same kind of strike twice or more in a row.
4) RECORD THE OTHER CONGA! (
)Regardless, your soundfont is way better than the conga built into my Ensoniq MR76 synth, which in general has great drum sounds. I'm sure folks will find excellent use for it.
Thanks again for sharing it!
Jeff
)-
- KVRAF
- 2844 posts since 1 Jan, 2003
I have to agree with Sascha here. I get very frustrated with sample sets that have all velocity layers normalized. Obviously there is no sound quality benefit (bit depth) if a sample is normalized after it's recorded. I also get frustrated when I can clearly hear more preamp noise and/or room level when someone has adjusted the recording levels to capture lower velocities. It strikes me as odd that when recording a real piano part in a real room, the engineer would never adjust the preamp level just for the quieter notes, bit depth be damned, but in sample libraries (especially drums) I can hear this un-natural effect often.
-
- KVRAF
- 13444 posts since 14 Nov, 2000 from Hannover / Germany
Thing with raising recording levels vs. normalizing vs. just recording at one level: What method would give the most convincing results?
I am not trying to give an answer but I'm really curious. Here's my thoughts:
- Normalizing samples: IMO most often a useless thing as your sampler of choice should just be able to do the job in its zone/group menu.
Would only be unsuitable in case the samples are REALLY low leveled, so the possible gain your zone/group menus would add still isn't enough. In all other cases I don't think it'll make too much sense as you may expect unwanted artifacts such as rounding errors and whatever, caused by normalizing (I'm by no means an expert, so someone correct me in case I'm wrong).
- Raising recording levels: Might eventually lead to similar unwanted artifacts as the normalizing thing, this time in the analog domain. Some preamps may tend to add more hiss when raising input gains. Of course, as the sample later on might not be played at a high volume, you may not notice these artifacts.
Still, unless you're using very high quality preamps there might be a difference in sound when raising input gains.
Of course, the advance of this method might be that a) good analog preamps might not add as much "errors" as normalizing and b) you would be able to use only low (or medium) velocity samples alone as well - without any gain adjustments, as the record level would be fine allready.
- Recording just away with one fixed input gain level (of course adjusted in a way that the loudest notes are captured as hot as possible): To me this seems to be the most "natural" thing. Consider you were doing, say, a piano recording - you wouldn't raise record levels for lower playing levels either during recording. Yes, on pp passages (or fading notes for that matter) one *may* notice the bit reduction - but I think with 24bit recordings done through quality equipment those errors should be pretty much barely noticeable.
So, why do it any different when sampling things? In the end, a sample set is nothing else but an authentic instrument recording, broken up into very small details (individual samples that is).
As said, I am not trying to give any answer or so - I simply don't know. How are the experienced sample heads handling things?
I am not trying to give an answer but I'm really curious. Here's my thoughts:
- Normalizing samples: IMO most often a useless thing as your sampler of choice should just be able to do the job in its zone/group menu.
Would only be unsuitable in case the samples are REALLY low leveled, so the possible gain your zone/group menus would add still isn't enough. In all other cases I don't think it'll make too much sense as you may expect unwanted artifacts such as rounding errors and whatever, caused by normalizing (I'm by no means an expert, so someone correct me in case I'm wrong).
- Raising recording levels: Might eventually lead to similar unwanted artifacts as the normalizing thing, this time in the analog domain. Some preamps may tend to add more hiss when raising input gains. Of course, as the sample later on might not be played at a high volume, you may not notice these artifacts.
Still, unless you're using very high quality preamps there might be a difference in sound when raising input gains.
Of course, the advance of this method might be that a) good analog preamps might not add as much "errors" as normalizing and b) you would be able to use only low (or medium) velocity samples alone as well - without any gain adjustments, as the record level would be fine allready.
- Recording just away with one fixed input gain level (of course adjusted in a way that the loudest notes are captured as hot as possible): To me this seems to be the most "natural" thing. Consider you were doing, say, a piano recording - you wouldn't raise record levels for lower playing levels either during recording. Yes, on pp passages (or fading notes for that matter) one *may* notice the bit reduction - but I think with 24bit recordings done through quality equipment those errors should be pretty much barely noticeable.
So, why do it any different when sampling things? In the end, a sample set is nothing else but an authentic instrument recording, broken up into very small details (individual samples that is).
As said, I am not trying to give any answer or so - I simply don't know. How are the experienced sample heads handling things?
Last edited by Sascha Franck on Mon Jan 03, 2005 5:55 pm, edited 2 times in total.
There are 3 kinds of people:
Those who can do maths and those who can't.
Those who can do maths and those who can't.
-
- KVRAF
- 7412 posts since 8 Feb, 2003 from London, UK
Much better post than what I was going to ask, Sascha Franck.
My question was, if I'm playing back a sample at the volume it was recorded at, how am I losing resolution more than when attenuating a sample to reduce its volume?
I mean, if I use 24 bit recording on a softly struck hit/note, and play that sample back at "full volume" when the MIDI velocity is "softly struck", then I'm getting the full 24 bit original recording, as recorded. If I amplify the sound in analogue and record that at 24 bits, then attenuate the signal to emulate "softly struck", surely I'm losing some resolution as I reduce the amplitude?
Surely it's better to avoid processing the original sound as much as possible, hence avoiding analogue noise or digital artifacts being introduced unnecessarily?
My question was, if I'm playing back a sample at the volume it was recorded at, how am I losing resolution more than when attenuating a sample to reduce its volume?
I mean, if I use 24 bit recording on a softly struck hit/note, and play that sample back at "full volume" when the MIDI velocity is "softly struck", then I'm getting the full 24 bit original recording, as recorded. If I amplify the sound in analogue and record that at 24 bits, then attenuate the signal to emulate "softly struck", surely I'm losing some resolution as I reduce the amplitude?
Surely it's better to avoid processing the original sound as much as possible, hence avoiding analogue noise or digital artifacts being introduced unnecessarily?
-
- KVRian
- 500 posts since 13 Oct, 2004 from Durham, NC USA
Coredelia, you have a good point about the danger of normalization causing different noise levels. I minimize that by always denoising my samples in CoolEdit.
Sasha, you have a point about naturalness. However, don't you think engineers would use a different recording level to record a piece that peaks at fff versus a piece that peaks at mf? Of course they would, if they're competent.
Soundfonts are compromises. Unless we want to use different sample sets for pieces with differing loudness, the best compromise is to carefully record the samples so that they're naturally normalized, taking care that the background noise level is reduced in all samples so that it's not audible.
Sasha, you have a point about naturalness. However, don't you think engineers would use a different recording level to record a piece that peaks at fff versus a piece that peaks at mf? Of course they would, if they're competent.
Soundfonts are compromises. Unless we want to use different sample sets for pieces with differing loudness, the best compromise is to carefully record the samples so that they're naturally normalized, taking care that the background noise level is reduced in all samples so that it's not audible.
This is incorrect. Soundfonts are limited to 16-bit samples, but I can record in 24 bits and increase by 36dB before hitting that problem. Furthermore, I recommend recording so that the samples seem normalized, rather than digitally normalizing.Cordelia wrote:Obviously there is no sound quality benefit (bit depth) if a sample is normalized after it's recorded.
-
- KVRAF
- 13444 posts since 14 Nov, 2000 from Hannover / Germany
Sure they would.learjeff wrote: Sasha, you have a point about naturalness. However, don't you think engineers would use a different recording level to record a piece that peaks at fff versus a piece that peaks at mf? Of course they would, if they're competent.
But in case of a sample set you just need to cover all occasions within a single patch - which, as you pointed out allready, will allways result in some sort of compromise.
There are 3 kinds of people:
Those who can do maths and those who can't.
Those who can do maths and those who can't.
-
- KVRian
- 500 posts since 13 Oct, 2004 from Durham, NC USA
Giving it a little more thought, I might take a different approach when recording in a live room with natural reverberation. Not having done that yet, it might well be that changing the gain between levels causes the room reverb to change unnaturally between the layers.
When recording "clean" -- either direct (like my Rhodes samples) or in a quiet recording studio, I stand by my advice above. But it might not hold in a live room.
When recording "clean" -- either direct (like my Rhodes samples) or in a quiet recording studio, I stand by my advice above. But it might not hold in a live room.
-
- KVRian
- 500 posts since 13 Oct, 2004 from Durham, NC USA
Right -- Sasha -- and that compromise doesn't include 9-bit samples at the tails of notes in an intimate passage!
Have you used Brad's soundfont yet? It's not feasible for a quiet conga part due to the quantization noise in his quietest samples, which peak as low as -30dB -- meaning the loudest bits are 11-bit samples! That's fine for a quiet note among loud ones. But it's not fine for a whole part that's to be quiet and intimate.
Have you used Brad's soundfont yet? It's not feasible for a quiet conga part due to the quantization noise in his quietest samples, which peak as low as -30dB -- meaning the loudest bits are 11-bit samples! That's fine for a quiet note among loud ones. But it's not fine for a whole part that's to be quiet and intimate.
-
- KVRian
- 500 posts since 13 Oct, 2004 from Durham, NC USA
I suppose we should take this to a new thread!
Sorry for railroading your thread, Brad. No good deed goes unpunished!
Sorry for railroading your thread, Brad. No good deed goes unpunished!
-
- KVRist
- Topic Starter
- 196 posts since 12 Mar, 2002 from Mountain View, CA
Wow, I never suspected I'd learn so much about sampling or trigger a debate! I'll take it as a compliment. 
Thanks for all the great tips. Jeff, thanks for pointing out the dead space in the beginnings of the samples. I'll take a few minutes to trim it and will upload "Brad's Conga 1.0.1" as soon as I'm done. Hopefully that will make the instrument more playable. I'll see if I can set up a time with my friend Brad to sample the other conga.
Boris
Thanks for all the great tips. Jeff, thanks for pointing out the dead space in the beginnings of the samples. I'll take a few minutes to trim it and will upload "Brad's Conga 1.0.1" as soon as I'm done. Hopefully that will make the instrument more playable. I'll see if I can set up a time with my friend Brad to sample the other conga.
Boris
-
- KVRAF
- 2844 posts since 1 Jan, 2003
"This is incorrect. Soundfonts are limited to 16-bit samples, but I can record in 24 bits and increase by 36dB before hitting that problem."
Ah, I hadn't thought of that!
Very interesting points, learjeff.
I guess you can tell from this thread, bburtin, that there are a lot of folks interested in your conga samples. I'll be checking back.
Ah, I hadn't thought of that!
Very interesting points, learjeff.
I guess you can tell from this thread, bburtin, that there are a lot of folks interested in your conga samples. I'll be checking back.