2009-11-11 14:16:59 speex echo canceller input span headroom
Rob Maris (GERMANY)
Due to several stability experiments with speex echo canceller, I have an impression that the fixed point DSP (including the blackfin specifics) may lose convergence status, resp. generate some extra noise on the output, when the signal input (microphone) and de reference input (speaker) come close to the 32768 limit - without clipping. Without thorough experimenting, I found that internal linearity of all processing is only guaranteed until approx. 2/3rd of max. signal span.
Anybody out there who can do a qualified statement about this? - thanks.