Advice on Bit Depth for recording

Farview said:
I'm pretty sure the -20 would rise more. But that's the sort of thing that happens with logarithmic scales.

Exactly. It would rise roughly in the ballpark of 3 times more because the dB is a logarithmic scale. The power requirements will be double or half every time you go up or down around 6.02 dB.

Farview said:
if you have a signal with 60db of dynamic range, it will have the same resolution peaking at -20dbfs as it does peaking at -1dbfs.

If that were the way the quantization scheme worked, the resulting audio would sound more like a telephone than a CD. A-law and m-law (non-linear) quantization works sort of like that, they give more power to the bottom of the scale. It has a few different effects. You'd probably have something like 8,000 cycles sample rate, 8 bits of depth and the digital noise would permeate the whole transmission. This is what Bell Labs did around WWII. They can get around the noise to a point by adding static to the signal to dither it and by noise shaping to move the noise out of the bandwidth the ear is more sensitive to. Also digital noise doesn't happen if there's no signal. It only happens because of rounding errors in the quantization.

The 8 kHz sample rate and 8 bit depth is to keep bit rates small enough to be feasable to transmit. The non-linear quantization creates a hyper-compression (LA 2A on steroids) effect. It's only designed for speech recognition and easy transmission. I've had people call me and say "Here, listen to this!" and they put the phone up to a speaker and I hear "Gwwwwqwwwvwvwvxxcvwwvwwwwzzx&&#!@#$%^&%$#" WTF?! It's useless for music.

For music recording and listening, PCM audio needs to have linear quantization to make the dynamics behave normally. The linear quantization needs to mimic what the amp needs to do to the analog signal. 6 dB per bit is handy this way. The ADC takes a snapshot of the signal every fraction of a second depending on where the sample rate is set, and measures the incoming voltage to quantize it to a signed interger value. 8 bits, 256 values to snap to. 16 bits, 65,536. 24 bits, 16,777,216. (<-- the majority of articles and tutorials I've read refer to this as "resolution") The "digital noise" or distortion happens when the resolution of low signals pretty much guarantees big rounding errors. If you start to notice that around -72 dBfs or so in 16 bit, it won't matter if you're peaking at -30 or -10. Change to 24 bit and you can add another 48 dB range, so where it distorts at -72 now becomes -120. Oops, now it's out of range so it's filtered out. Gone. No dither required prior to requantizing.

See what I mean?

Aside from all this stuff, any relation to Joe?