Recognizing speech at 1bit quantise depth? i found on german wikipedia an audio example of 1 bit depth quantising, where the speech still can be recognized. how is it possible if at 1 bit depth we have just two values: "signal" and "no-signal"?. here is the examle: https://upload.wikimedia.org/wikipedia/commons/4/43/Ampl1rp.ogg
 A: A 1 bit depth quantised signal still contains more than one bit of information. The signal level varies from moment to moment, and this provides extra information. 
In the case of speech we tend to recognise the rhythmicity and structure as speech even if we cannot make out individual words. Some formant sounds may be recognisable if they get the signal stream to flip from 0 to 1 and back again at the same frequency: we tend to recognise voice sounds by their peak formant frequencies, so again this helps recognizing the signal as speech.
A: 1-bit quantization involves sign-only sampling, which can be done at fantastic rates, well over 10^9 samples per second.  For purposes of speech recognition, frequencies over 6 kHz are irrelevant.  If a low-frequency signal plus wideband noise is subjected to 1-bit quantization, nonlinear phenomena can sometimes interfere with extraction, but they don’t always do that.  Consider the following cases: 
(1) A low-frequency signal of high amplitude plus wideband (think white) noise of low amplitude.  The noise can throw off the recorded sign of the signal whenever the signal crosses zero, but this sort of error averages to zero.  A bigger problem is that sinusoidal signals get converted to square waves, so there are nasty intermods -- beats and harmonics.  In the simple case of square waves, the harmonics contain only 20% of the power.  Ratios of amplitudes will be distorted in mixtures of overtones, possibly preventing the identification of vowels distinguished by ratios of overtones in the formant regions of the spectrum.  
(2) A low-frequency signal of low amplitude plus wideband gaussian noise of high amplitude.  The weak signal will register only when the noise happens to be smaller, but that happens often enough to allow extraction with minimal distortion by intermods.  There will be 2 dB of nonlinear suppression relative to many-bit quantization.  Ultimately, a high sampling rate saves the day.  Averaging N uncorrelated samples will enhance the signal-to-noise ratio by a factor of N.
