bibilolo from recent 64 kbps test is a very hard sample for wavPack.
Original is sample 3 from Gabriel's samples for an 48 kbps AAC test: http://www.mp3-tech.org/tests/aac_48/samples/ (thanks AlexB for pointing me there).

Sample is rather quiet and volume should not be turned too high in order to preserve a realistic listening condition. But even then hiss and/or distortion are very obvious.

Even with 500 kbps fast mode is pretty inadequate. Normal or (better) -h mode improves things. A positive noise shift has a good influence in this case too, as does a higher -x setting. Bitrate chosen by wavPack lossy is lower than nominal (a rare event) and this is what makes the quality so bad.

With my resampling to 32 kHz and using 380 kbps (roughly the same per sample accuracy as 500 kbps with 44.1 kHz samples) things are even worse. Quality difference between normal/high and fast mode is bigger than in the 44.1 kHz case.

With this in mind I see my resampling procedure in a different light. With 44.1 kHz may be there's more space for inaudible or hardly audible added noise especially when shifted up.
For a compromise between usually overkill quality and acceptable quality with a sample like this I'll use -hb450x4s0.8 for my new encodings.

Its like the martenotwaves samples of guruboolez (dentist drill). Problem for the format using standard compression parameters. For these use normal mode + x4 or high modes, But even plain 400k normal mode x1 gave good results.

I did some compression tests with various modes:

fx = 48.49%
x = 52.10%
hx= 56.50%

x4 = 57.54 %
hx3 =57.19 %
hx4 =59.86 %

perhaps -b400 x4 or hx is good space / quality compromise even on harder samples like this.

For people who like to try out bibilolo: Second 17.2-19.2 for instance is a remarkable spot.

I also tried 3.xx experimental auto noise shaping, and this makes things extremely bad.

In case auto noise shaping cannot be improved compared to the behavior of this experimental version I think quality control is what is neeeded most with wavPack lossy. Such a pity that wavPack lossy which usually is great in the lower 3xx kbps range currently doesn't deal well with such cases (which are rare but very ugly when they happen).

@shadowking:
a) Do you have a link to martenotwaves?
b) Where to put the sweetspot in a compromise is always a matter of taste.
With usually achieved quality in primary focus while still respecting weird samples like these to a certain exent -b400x4 maybe fine fine (though on sec. 17.2-19.2 of this sample it's bad). With quality in primary focus also with these weird samples while still respecting the fact that it's usually overkill, a higher quality setting like
-hb450x4 is welcome, while with highest quality in focus for weirdest samples even a higher quality setting is necessary.

Replaygain is +15 db !.. Could only be an issue with track gain. I won't listen that high.

BTW - Dualstream bitrate is 460k for fast mode and 400 for normal (quality 3) which is a good thing here.

This could be the 'achilis heel' of ABR as Guruboolez said. We don't know it as well because more people are using VBR. I have seen it before on mp3 and aac - bitrate isn't high enough on ABR which heavy transient stuff and bitrate efficiency is always lower than VBR.

Yes, it's a quiet sample, and volume should be kept in a position as with 'normal' loudness tracks.
There's always the risk of misinterpreting things with quiet track excerpts as we have a tendency to turn up volume. I did this when I heard bibilolo for the first time.

But also with volume at a reasonable level this is a hard sample. Did you try sec. 17.2-19.2 (no replaygain of course) using -b400x4?

Nice to hear that Optimfrog VBR is fine. BTW I wouldn't call wavPack lossy an ABR method as this resembles Lame's ABR. As far as I can see wavPack lossy doesn't have any real quality control. It just encodes the prediction error with an accuracy given by the chosen bitrate (or given bit per sampe accuracy), and has some inertia properties in this context. It's very respectable that this usually works very well even at pretty moderate bitrate, but it's not a method I'd call ABR. Encoding the prediction error with a certain accuracy makes sure that resulting bitrate is pretty constant, and the fluctuations come in because of the inertia properties (which as a tendency have a good influence on quality but not in a controlled way). I wouldn't call that ABR as from Lame understanding ABR means bitrate variation depending on local bitrate demand with respect to quality (under the side condition of a target bitrate). Lame's ABR is a variable bitrate method but with a different quality control than VBR which has its pros and cons.
IIRC David Bryant has called wavPack lossy a CBR method (which when compared to mp3 CBR is also not totally comparable as due to bit reservoir even mp3 CBR does a [restricted] bitrate variation according to local demands).