lossyWAV Development
I just finished my abx test. First I followed your suggestion and used -7 -autoshape -nts 20 -snr 14. This setting yields 309 kbps for my regular set (quite a bit higher than plain -7) and 355 kbps for my problem set (a bit low for problem samples). Hiss is pretty audible with bruhns (for instance sec. 9.3-10.2), but it's audible also with bibilolo (sec. 4.3-5.5) and badvilbel (sec. 5.9-7.2). There's also a slight inaccuracy with Atemlied (sec. 9.3-10.1) which is best audible at moderately high listening volume. I didn't test a lot more samples then those mentioned because to me this is not adequate quality for an average of 309 kbps. The hiss (and the inaccuracy) isn't really annoying though, and I listened to some regular music (carefully but without abxing), and was content with it. Anyway looking at codecs like vorbis (I just tested the new Aoyumi version, and quality is great even at -q4 [~130 kbps] I personally don't like my abx result at a bitrate of ~310 kbps. I redid the test using plain -7 autoshape which yields 325/384 kbps for my regular/problem test set. bruhns 9.3-10.2 is better now to me, though still quite audible. Same goes for bibilolo. I didn't test more samples as I personally am not content with this as well. -6 -autoshape yields 337/404 kbps. bibilolo is ok now, but when abxing bruhns I found added hiss already at sec. 2.3-4.4. I skipped -5 and went directly to -4 -autoshape as from my last test I know plain -4 is transparent for me with the samples tested. -4 -autoshape yields 369/450 kbps. With this average bitrate for the problem set chances are good that everything is alright now. bruhns at sec. 9.3-10.2 however still isn't perfect though quite acceptable. Summing it up to me this isn't a good result for using -autoshape though it's a matter of taste whether or not one is willing to accept added hiss which seems to be the major issue when using -autoshape with low bitrate settings. Maybe a variant of autoshape is more succesfull: make the frequency up shift of noise also depend on the degree to which there's energy in the input signal's 2 highest frequency zones (that is the range from ~8.2 kHz up). Do you like to try that, Nick? Ok, I will try again to implement the RMS variability approach that I was trying (but didn't release). Another approach would be to make the variability of the shaping non-linear with respect to bits-to-remove. At present it increases at 1/13 per bit to remove, i.e. 0=0; 1=1/13; 2=2/13; etc; 12=12/13; 13=13/13. If I was to change this from linear to some power, say for example shaping_factor = 1-((13-bits-to-remove)/13)^n then things may change. Again, the noise shaping function itself is totally fixed, all the autoshape function is vary how much to apply. It doesn't change the frequency to which the noise is shifted. Think of it as -shaping 0 = pure white noise; -shaping 1 = fully shaped noise; -shaping <n> = something in between. I'll get back to the "drawing board" with the autoshape function and post v0.9.1 soon. As an aside, I have found that TCPMP for my iPAQ plays FLAC *much* better (better = less cpu usage and more accurate output) than GSPlayer / gspflac.dll. In particular dithernoisetest would exhibit some harmonics using GSPlayer which don't exist in TCPMP. TCPMP v0.72 RC1 is still available, google is your friend....