[Xiph-Advocacy] example of lossless fade out

[Tor-Einar Jarnbjo]

xiphmont at xiph.org schrieb:
> You're conflating the floor and the window.  Long frames are still
> multiplied by the floor, even if the window is a different shape.
>   

Huh? What I ment is that in the transition between a short and a long 
frame, the window function applied to the long frame has a section, in 
which only samples from the long frame are copied to the PCM output. 
Explained on the example in the Vorbis specification:

http://xiph.org/vorbis/doc/Vorbis_I_spec.html (Figure 3: overlap of a 
long and a short window)

The samples in the period from the middle of the red frame until the 
start of the green frame (384 samples) are only taken from the red 
(long) frame and not mixed with the previous or following frame. If each 
frame is attenuated completely with a fixed factor and the attenuation 
is increased for each frame to produce a fading effect (of course taking 
into consideration that the attenuation must be increased more per frame 
for long frames than for short frames), this section will be output with 
a "break" in the fading envelope. I just estimated the few ms I 
mentioned in the last message, with these parameters it's actually less 
than 1 ms, making it even less probable that this can actually be heard.

As Gregory mentioned, changing the floor will of course require the 
entire floor to be reencoded and not just "bit fiddling".

Tor