What is best case s/n ratio in audio processed in 32 bit floating point? -- Dirk http://www.transcendence.me.uk/ - Transcendence UK http://www.theconsensus.org/ - A UK political party http://www.blogtalkradio.com/onetribe - Occult Talk Show
Best Case Noise
Started by ●February 15, 2010
Reply by ●February 15, 20102010-02-15
On 15 Feb, 18:32, Dirk Bruere at NeoPax <dirk.bru...@gmail.com> wrote:> What is best case s/n ratio in audio processed in 32 bit floating point?Yes and no. 'Yes' if you are talking pilosophically, in the case where the signal is exactly representable by the floating-point values at the sampling instants. 'No' if you are talking practically, since there will always be *some* sort of noise present, e.g. by internal noise in the sampling device, or quantization. Rune
Reply by ●February 15, 20102010-02-15
Dirk Bruere at NeoPax wrote:> What is best case s/n ratio in audio processed in 32 bit floating point?Very-very-very best case? Somewhat 1500 dB. Are you happy now? VLV
Reply by ●February 15, 20102010-02-15
Vladimir Vassilevsky wrote:> > > Dirk Bruere at NeoPax wrote: > >> What is best case s/n ratio in audio processed in 32 bit floating point? > > Very-very-very best case? Somewhat 1500 dB. Are you happy now?Moron. -- Dirk http://www.transcendence.me.uk/ - Transcendence UK http://www.theconsensus.org/ - A UK political party http://www.blogtalkradio.com/onetribe - Occult Talk Show
Reply by ●February 15, 20102010-02-15
Rune Allnor wrote:> On 15 Feb, 18:32, Dirk Bruere at NeoPax <dirk.bru...@gmail.com> wrote: >> What is best case s/n ratio in audio processed in 32 bit floating point? > > Yes and no. > > 'Yes' if you are talking pilosophically, in the case where the > signal is exactly representable by the floating-point values at > the sampling instants.Yes. The irreducible noise that comes with the arithmetic. I was wondering whether it is better than the theoretical 23 bit precision (23 + sign + 8 exponent) -- Dirk http://www.transcendence.me.uk/ - Transcendence UK http://www.theconsensus.org/ - A UK political party http://www.blogtalkradio.com/onetribe - Occult Talk Show
Reply by ●February 15, 20102010-02-15
Dirk Bruere at NeoPax wrote:> Rune Allnor wrote: > >> On 15 Feb, 18:32, Dirk Bruere at NeoPax <dirk.bru...@gmail.com> wrote: >> >>> What is best case s/n ratio in audio processed in 32 bit floating point? >> >> >> Yes and no. >> >> 'Yes' if you are talking pilosophically, in the case where the >> signal is exactly representable by the floating-point values at >> the sampling instants. > > > Yes. > The irreducible noise that comes with the arithmetic. > I was wondering whether it is better than the theoretical 23 bit > precision (23 + sign + 8 exponent)For the same number of significant bits, floating point has ~3dB better SNR then the fixed point normalized to maximum value. Feeling better now? Stupident :))))) VLV
Reply by ●February 15, 20102010-02-15
Vladimir Vassilevsky wrote:> > > Dirk Bruere at NeoPax wrote: > >> Rune Allnor wrote: >> >>> On 15 Feb, 18:32, Dirk Bruere at NeoPax <dirk.bru...@gmail.com> wrote: >>> >>>> What is best case s/n ratio in audio processed in 32 bit floating >>>> point? >>> >>> >>> Yes and no. >>> >>> 'Yes' if you are talking pilosophically, in the case where the >>> signal is exactly representable by the floating-point values at >>> the sampling instants. >> >> >> Yes. >> The irreducible noise that comes with the arithmetic. >> I was wondering whether it is better than the theoretical 23 bit >> precision (23 + sign + 8 exponent) > > For the same number of significant bits, floating point has ~3dB better > SNR then the fixed point normalized to maximum value. > > Feeling better now? Stupident :)))))Well dumbo - why use float instead of 24 bit int then? You think that maybe... just maybe... there's something a bit more to it than that when dealing with large signals and/or cascaded filters? -- Dirk http://www.transcendence.me.uk/ - Transcendence UK http://www.theconsensus.org/ - A UK political party http://www.blogtalkradio.com/onetribe - Occult Talk Show
Reply by ●February 16, 20102010-02-16
Dirk wrote:>What is best case s/n ratio in audio processed in 32 bit floating point?If you just meant "represented" in 32-bit FP, that's another question. After all, your audio signal *could* be zero. Not very likely. You say "processed", which is not really enough information. But FP epsilon might provide one bound, if you assume very small noise (quantization, perhaps, if your signal is synthesized, rather than measured). I can add 2^-23 to 1 in IEEE single-precision float, and still tell the difference from 1, but I can't do that with 2^-24. This seems reasonable, since there's a 23-bit mantissa, plus an implied "1" in that format (barring denormals). From there, you might get a crude bound on SNR after processing, though it will generally be worse than this if your algorithm is not designed carefully. Another way to guess at a lower bound would be to run a carefully-constructed algorithm against it. FFTW serves a dual-purpose here, in also providing a spectrum. So you can generate a sine for a length a multiple of its period (so as to not need a window), FFT in single-precision, and plot the magnitude spectrum. That should be pretty close to the limit, unless I've missed a step in there...
Reply by ●February 16, 20102010-02-16
Michael Plante wrote:> Dirk wrote: >> What is best case s/n ratio in audio processed in 32 bit floating point? > > If you just meant "represented" in 32-bit FP, that's another question. > After all, your audio signal *could* be zero. Not very likely. > > You say "processed", which is not really enough information. But FP > epsilon might provide one bound, if you assume very small noise > (quantization, perhaps, if your signal is synthesized, rather than > measured). I can add 2^-23 to 1 in IEEE single-precision float, and still > tell the difference from 1, but I can't do that with 2^-24. This seems > reasonable, since there's a 23-bit mantissa, plus an implied "1" in that > format (barring denormals). > > From there, you might get a crude bound on SNR after processing, though it > will generally be worse than this if your algorithm is not designed > carefully. > > Another way to guess at a lower bound would be to run a > carefully-constructed algorithm against it. FFTW serves a dual-purpose > here, in also providing a spectrum. So you can generate a sine for a > length a multiple of its period (so as to not need a window), FFT in > single-precision, and plot the magnitude spectrum. That should be pretty > close to the limit, unless I've missed a step in there...Probably best to stick to 64 bit float then. Only the i/o has to be 24 bits. -- Dirk http://www.transcendence.me.uk/ - Transcendence UK http://www.theconsensus.org/ - A UK political party http://www.blogtalkradio.com/onetribe - Occult Talk Show
Reply by ●February 16, 20102010-02-16
>Vladimir Vassilevsky wrote: >> >> >> Dirk Bruere at NeoPax wrote: >> >>> Rune Allnor wrote: >>> >>>> On 15 Feb, 18:32, Dirk Bruere at NeoPax <dirk.bru...@gmail.com>wrote:>>>> >>>>> What is best case s/n ratio in audio processed in 32 bit floating >>>>> point? >>>> >>>> >>>> Yes and no. >>>> >>>> 'Yes' if you are talking pilosophically, in the case where the >>>> signal is exactly representable by the floating-point values at >>>> the sampling instants. >>> >>> >>> Yes. >>> The irreducible noise that comes with the arithmetic. >>> I was wondering whether it is better than the theoretical 23 bit >>> precision (23 + sign + 8 exponent) >> >> For the same number of significant bits, floating point has ~3dB better>> SNR then the fixed point normalized to maximum value. >> >> Feeling better now? Stupident :))))) > >Well dumbo - why use float instead of 24 bit int then? >You think that maybe... just maybe... there's something a bit more to it >than that when dealing with large signals and/or cascaded filters? > >-- >Dirk >Sometimes I read these discussions just for the entertainment value... Jacob
