Test data

[viciious]

Hi!

Thanks for your work on adpcm-xq! It's greatly appreciated :)

If anything, it'd be nice to have some test data corpus that would accompany the utility, along with some SnR measurements. Not strictly necessary, but desirable :)

Regards,
Victor

[dbry]

Hi Victor! Sorry I didn't answer earlier. I agree that such a test would be good for comparing adpcm-xq to other algorithms. I have some standard speech files that might be perfect. Unfortunately I am currently swamped with existing projects, so this will have to come some day I get bored... :)

One important thing is that for an apples-to-apple SnR comparison, the dynamic noise shaping feature in adpcm-xq must be turned off. The reason is that noise-shaping increases the noise level (by definition), but hopefully decreases the audibility of such noise (which is, of course, more difficult to quantitatively measure).

[donnaken15]

I have some short sound files (4-25kb) of sort of basic waveforms where it appears to prevent encoding noise only on a few sounds but not others

[dbry]

I am currently working on this program and would be happy to look at any samples that you're having trouble with.

BTW, in the newest versions I added the ability to do noise measurements when encoding (not a ratio but absolute noise level).

[donnaken15]

Here's what I tried this utility on:
samples.zip
the files are suffixed with their bit depth, and for ones that are ADPCM, I only included the ones that had barely audible noise or difference
I compiled the latest commit myself two days ago and got some weird thing on one sound where higher frequency noise appeared to be reduced only a little the more I put -s towards -1.0 compared to using something like -f but as if it somehow distributed and multiplied that to low frequencies like below 300 Hz, which is not meant to be part of the sound

[dbry]

Thanks for the samples! These behave a little differently than the mostly music and speech samples that I normally use so I'll keep them. I even have an upcoming tweak to the non-exhaustive mode because I noticed on some of these the -x option reduces noise (which is not the case with my other samples).

Your results with the -s option below 0.0 actually make sense. It reduces the high-frequency noise at the expense of more noise down below. The idea is that most of the audio power occurs at lower frequencies (at least at 44/48 kHz SR) and so it would normally mask the noise. Of course, if you have a sample with nothing below 300 Hz then the noise would become more audible down there (although that might be less objectionable than hiss up above). In any event, I would not go below -0.5 for the -s option unless there was lots of LF audio to mask the noise and the HF hiss was audible.

I will be updating the dynamic noise shaping very soon and I also limit the negative shaping to -0.5 in that mode. With any luck that mode will eliminate the need to fiddle with the -s option.