Hi,
OK, basically it appears to be an overflow of the decoded signal, but I
can't reproduce it with speexenc/speexdec. There's an overflow occurring
instead of what should be saturation. Is it possible you're using
speex_decode() and converting from float to short without first checking
for overflows? Normally speex_decode_int() should do it for you, but if
you use the float version (speex_decode), then you need to do it
yourself. The reason you only see the problem in beta2 and not beta1 is
probably more of an accident than anything else.
Cheers,
Jean-Marc
P.S. Oh, and BTW, 22 kHz isn't a recommended sampling rate.
Aviv Tamar wrote:> Hi,
> I attached the original voice file, and the decoded voice files on the two
> last versions. beta1 is clearly better.
> (See attached file: bg5_sab_1368.wav)
> (See attached file: bg5_sab_1368_beta1.wav)
> (See attached file: bg5_sab_1368_beta2.wav)
> Also, I found that my problem is in the decoder (I used encoders and
> decoders from different version to isolate the problem).
> I am using floating point.
> Thanks,
> Aviv
>
