I just typed a lengthy question and my browser crashed so I’m going to paraphrase it…
I record in 24-bit solely so I can enjoy the headroom. My target format is always 16-bit for CD.
Most books and articles I’ve read suggest setting levels (on the original 24-bit master) to around -12dB for this kind of application. But my question is why you can’t go much lower still? In theory you could peak at -48dB and still lose no data once you’ve re-quantised. Am I correct in saying the reason this would be poor practice is because of the analogue noise floor? In other words the bottom 3 or 4 bits are likely just (analogue induced) noise, so I would want my recorded peaks to be at least 16-bits-worth (96dB) higher than this? Thank you.
If Audacity is set to record in 32-bit float (the default), then it will use the highest format it can get.
32-bit float is the recommended working format as it protects against clipping at 0 dB, and Audacity works internally in 32-bit float, so it avoids unnecessary format conversions if you work entirely in 32-bit float.
If the “host” is set to MME, or direct sound (in the device toolbar), then you will get 16-bit audio data, regardless of other settings in Audacity. This is because the interface between Audacity and the computer sound system (“PortAudio”) only supports 16-bit for MME (and Direct Sound is emulated on Vista and later).
You “should” be able to get 24-bit data from your sound card if:
Audacity is set to record either 24-bit or 32-bit float (32-bit float is the strongly recommended option).
This depends on the sound card, but even state of the art sound cards are just noise in the bottom couple of bits. Cheaper sound cards are likely to have more digital noise than state of the art. Nevertheless, a reasonable quality 24-bit audio device should still give you more dynamic range than 16-bit.
Common practice in recording studios is to aim for around -20 dB. This allows plenty of leeway in both directions and saves time setting up (studio time is expensive).
Basically I agree and I think people worry too much about recording levels. As long as you’re not clipping, you’re usually OK. (Even at 16-bits the analog noise usually is what limits the usable dynamic range.)
Am I correct in saying the reason this would be poor practice is because of the analogue noise floor? In other words the bottom 3 or 4 bits are likely just (analogue induced) noise
Typically you’re attenuating the analog signal and noise together, so the analog signal-to-noise ratio is not affected. If you’re recording at a low level because you’re far from the mic, or because you have a weak analog signal, then yes, your signal-to-noise ratio will be degraded.
There is quantization noise at low levels and you can hear quantization noise at 8-bits.
I have read that most 24-bit ADCs are only accurate to about 20-bits, but I’m not sure what that means… If those lower bits are random, that’s noise and it could be a big problem at very-low levels.
Most books and articles I’ve read suggest setting levels (on the original 24-bit master) to around -12dB for this kind of application.
It depends on what you’re recording. If you’re recording “live” you need to leave room for unexpected peaks. If you are digitizing records or tapes then the peaks are more-predictable and you can record hotter. The headroom is for ONLY unexpected peaks. If your highest peak turns-out to be -12dB, you didn’t really need the headroom. But if you’re recording at -12dB and after recording you find a 0dB peak, the data is probably clipped and you needed more headroom.
I’ve suspected that the tradition of pros recording at -18dB came from when Pro Tools was integer based. You needed “extra headroom” for mixing & effects. With floating-point processing we can “temporarily” go over 0dB without clipping.