Typical signal to noise ratio of ADPCM Options

[Neuron]

So, I have decided to return to this forum (despite my little conflict with greynol) and I have kind of a weird question. I would like to know what is the rough SNR of a 4-bit ADPCM recording (IMA and Microsoft). For those curious why I am asking this, well, I like ripping music and sound effects from old games, and I would like know the technical side as well.

[Neuron]

I did an Audacity convert to mono + match both volumes using Amplify + invert one of the track and Mix and render way of subtracting one track from the other. I got a "noise signal", peaking only 11.3 dB below the top. The weird thing is, the quieter parts of the song (not silence, only 9-10 dB below maximum) had "noise" as quiet as -45 dB while the louder parts went up to 11.3 dB. This is not all that is strange, because the "noise" left by the subtraction is actually a recognisable signal of the music, although very noisy (I will post a sample in the uploads section). By contrast, the dithered 8-bit signal is very audibly noisy, but it leaves a normal noise signal at -35.3 dB (which is a bit weird as well as 8-bit sound should have a 48 dB SNR and dithered 42 dB as I used only 1-bits of triangular dither, plus, the original signal is slightly present in the noise as well). The undithered 8-bit signal produces normal random noise at -48.2 dB as expected. Both dithered and undithered 8-bit signal have obvious, strong noise in them, unlike the ADPCM files.

So, how to make sense of this? And why every difference signal except for the non-dithered 8-bit one has recognisable music it it?

[saratoga]

I did an Audacity convert to mono + match both volumes using Amplify + invert one of the track and Mix and render way of subtracting one track from the other. I got a "noise signal", peaking only 11.3 dB below the top. The weird thing is, the quieter parts of the song (not silence, only 9-10 dB below maximum) had "noise" as quiet as -45 dB while the louder parts went up to 11.3 dB. This is not all that is strange, because the "noise" left by the subtraction is actually a recognisable signal of the music, although very noisy (I will post a sample in the uploads section). By contrast, the dithered 8-bit signal is very audibly noisy, but it leaves a normal noise signal at -35.3 dB (which is a bit weird as well as 8-bit sound should have a 48 dB SNR and dithered 42 dB as I used only 1-bits of triangular dither, plus, the original signal is slightly present in the noise as well). The undithered 8-bit signal produces normal random noise at -48.2 dB as expected. Both dithered and undithered 8-bit signal have obvious, strong noise in them, unlike the ADPCM files.

So, how to make sense of this? And why every difference signal except for the non-dithered 8-bit one has recognisable music it it?

Have you looked at the wikipedia page on ADPCM? What you're describing is basically the difference between PCM and ADPCM:

The default signal compression encoding on a DS0 is either μ-law (mu-law) PCM (North America and Japan) or A-law PCM (Europe and most of the rest of the world). These are logarithmic compression systems where a 13 or 14 bit linear PCM sample number is mapped into an 8 bit value.

So it's compression involves taking a 14 bit signal (or some other value depending on the flavor of ADPCM) and spacing the levels on a log scale. So if the signal is very quiet, the levels are close together, and the error is smaller. If the signal is larger, the error is larger. If the error was constant, you'd have regular PCM.

[Neuron]

So it's compression involves taking a 14 bit signal (or some other value depending on the flavor of ADPCM) and spacing the levels on a log scale. So if the signal is very quiet, the levels are close together, and the error is smaller. If the signal is larger, the error is larger. If the error was constant, you'd have regular PCM.

So if I want a good ADPCM sample I should lower the volume of the uncompressed file first? If so, by how much?

[saratoga]

So it's compression involves taking a 14 bit signal (or some other value depending on the flavor of ADPCM) and spacing the levels on a log scale. So if the signal is very quiet, the levels are close together, and the error is smaller. If the signal is larger, the error is larger. If the error was constant, you'd have regular PCM.

So if I want a good ADPCM sample I should lower the volume of the uncompressed file first? If so, by how much?

Reducing the signal does reduce the error, but since SNR is the ratio of signal to noise . . .