I’m about to go through all the factory SFZs (as well as my own sfz packages) and level the volumes. I’d like input, though: what’s the target?
Here’s my take:
polyphonic instruments: play a 4-note chord around Middle C
monophonic: play Middle C
Use 96 (75%) for velocity, CC7 (volume), and CC11 (expression).
Use LUFSMeter to measure, and set sfz global_volume to get target LUFS value.
Make sure peak level stays below -10dBFS.
Use a MIDI file to play consistently. The chord or note plays repeatedly with keys held down most of the time.
What should the target LUFS level be?
LUFS is fairly close to RMS, but is adjusted to reflect human loudness perception. It’s also not far off from VU, if you’re old enough (like me) to have an idea what VU levels mean. If you don’t have any sense of these units, feel free to answer in dBFS and I’ll do my best to compensate.
If the answer is different for different kinds of instruments, just list the ones you have opinions about.
There might be a special case for non-velocity-sensitive sounds like organs. I’ll have to give that some thought and fiddling.
And finally, percussion is definitely a special case (not melodic percussion like marimba, but drums & such.) I plan to leave them alone unless anyone has specific requests. Too far outside my wheelhouse.
Please voice your opinions on targets, or methods, or anything I didn’t mention that should be considered.
In the broadcast industry we equated “VU” with “virtually useless” because it didn’t give us a useful measure of peak programme levels. We used a peak meter with weighted response and slugged release. The meters on zynthian are designed like this and are calibrated so I wonder why you haven’t suggested using them, rather than another device.
I would suggest nominal peak programme of a large chord to hot the amber section, below the red. I am away from home now, but IIRC the red is at -10dbFS and the orange is at -18dbFS. We used floating point internally so as not to clip but clipping will occur when the signal hits the output. Percussion is a challenge. There will be transient peaks the are very loud within a quieter sound. You can often allow these to overload because they are spikes that you may not detect distortion but you really want those peaks to be below 0dbFS.
Zynthian meters also have peak hold and overload indication which helps with setting levels.
VU was virtually useless during recording and broadcast because they don’t prevent exceeding the capacity of the medium. They were more useful than peak meters for estimating loudness, though.
The best case is two indicators, one for peaks and another for volume (LUFS), and the peak indicators should leave a high-water mark that slowly sinks back down. If Zynthian uses the continuous mark for something approximating volume and the drifter for actual peak, that’s great.
Note that LUFS are compensated by the K filter too, which approximates human hearing response to frequencies.
Peak meters are wrong for adjusting volume, because they don’t correlate to volume very well at all. LUFS was defined for exactly that purpose; perceived volume.
I’ll be interested to see how Zynthian meters compare to actual peak and LUFS meters.
I like your suggestion for targets. Good to leave lots of headroom.
I would expand the chord, so as to better capture the amplitude excitation of the combined highs and lows. I would suggest a four-note C major chord above middle C (right hand) with an octaved C2+C3 bass below (assuming C4 is middle C).
Should we make it 100 for both velocity and CC? (considering that it reflects most Midi insert standard levels in DAWs).
On radio, everything drowns in a swamp of unified overcompressed loudness.
On TV and streaming services, dialogue is almost drowned out by constant, deafening pad sounds.
Thanks to psychoacoustic sound modeling!
They are all trying to save on sound engineer costs through automation. At what price?
Nowadays, dynamic range no longer matters. It just has to be loud enough!
Remember, the CD-DA gave us 96dB of dynamics, a breakthough!
Instead dynamics are mow compressed down to 16bB or less and then peaked out maximum level. As long as it is loud enough!
The end might be a true binary 1-bit sound.
Maximum volume, maximum compression, minimal dynamic range. Simple and inexpensive.
And it wouldn’t sound much worse than many of today’s audio.
The production of (good / quality) radio cares immensely about levels and dynamic range. The studio output is well balanced. It then feeds processors that adjust the sound to meet the station’s requirements (and the technology of broadcast). This might be multiband compression to stand out on the band. It may be gentle compression and limiting to provide wide dynamic range for listening in an armchair. It may be more aggressive compression at drive-time to compensate for noise from the commute. Our job is to create the tools to allow the composer / performer to create their masterpieces. Those same artists can then cry when they hear what it sounds like on there FM radios!
My point is that we should aim for levels that naturally sit in the mix but can be adjusted by the performer. We shouldn’t create an oboe that is quieter than a trumpet because that is how they sound in real life. Both should peak to a sensible level so that they can be mixed without excessive loss of quality. (I tend to normalise samples so that they have maximum dynamic range allowing them to be attenuated with minimal impact on fidelity.) Quiet samples can introduce noise when boosted… but we may already be constrained by the existing samples.
(I may have mentioned working in broadcasting for several decades, and specifically radio.)
I suspect you never worked for Clear Channel. I knew an audio engineer who did. He cared about all that stuff, but CC’s goals were a bit different. Anyway, that doesn’t undermine anything you said. You and @fussl are saying opposite things, but I believe that you’re both correct.
