As a follow-up to some discussion around P#6159, there is a concern that
adding many new entries in the language file might make the voice file
grow beyond the buffer space available on the Archos.

Depending on speaking rate and compression parameters, I have generated
voice files with sizes varying between ~1M and ~1.5M, and that's including
all the new voice entries in all the patches I've submitted so far.

So, do we actually have a problem, or not?

I guess not for a little while longer, since AFAICT we can still compress
the file small enough.

Still, I've gathered some stats to try to understand this:
726 entries in my own .lang file,
169 have no voice text, so
557 non-empty voice clips.

Looking at all voice texts, the texts are composed of 1050 total words,
with 531 distinct words.

198 words are repeated, for a total of 519 repetitions.
117 words of 5chars or more are repeated, 290 repetitions.
35 words of 5chars or more are repeated 4times or more, 190repetitions.

Worst offenders:
Word Count
frequency 6
backlight 8
bookmark 8
equalizer 9
settings 12
directory 13
playlist 17

The average length of repeated words is pretty much the same as that for
non-repeated words: 5.51chars vs 5.58chars.

If we consider only words that are 5chars long or more:
290 repetitions out of 1050 total words = 27.6%.
If we consider all words, including short ones that are only one syllable:
519 repetitions out of 1050 total words = 49.4%

So if we could implement something that reuses individual word clips and
concatenates them, we could save anywhere between 25% to 50% on the voice
file size. I'd rather not get into that... (and my own player doesn't need
it ;-) ) but that's an option to look at if we need to. Of course, it's
likely to insert small pauses between words, making rockbox sound like
those cheap automated phone services...

-- 
Stéphane Doyon
<s.doyon_at_videotron.ca>
http://pages.infinit.net/sdoyon/