Re: [Audacity-devel] Fwd: SoX resampling

[Roger D. <rb...@cs...>]

On 9/23/12 8:05 PM, Steve the Fiddle wrote:
> Also, setting the range of our Spectrogram out to 180dB, to match the
> >http://src.infinitewave.ca/  analysis, both their float test signal and a
> >generated signal in Audacity don't look great.  Is that a limitation/feature
> >of our Spectrogram code, a limitation of the 'float' nature of the signals,
> >or something else?
> >
Each bit doubles the signal to noise ratio, which is an increase of 
about 6dB, so 180dB implies about 30 bits of precision, but 32-bit 
floats have at most 25 bits (23 actual bits, 1 hidden bit, and 1 sign 
bit). FFTs are not accurate to the least significant bit, so there's 
some loss of precision there too.