Quote Stoney:

---

But is it correct that only the frequency below this threshold can be conveyed with any accuracy? The amplitude would be whatever the dither happens to be - no?

---

Depends if you are talking about what the human ear/brain perceives, or what a broadband audio test set would measure.

The dither signal is a random noise, where as the audio signal is, to a degree, recognisable and predictable. As such, the human ear/brain is able to ignore the random noise and determine with accuracy both the frequency and amplitude of the signal below the noise floor.

What a broadband test set would measure is, essentially, just the dither noise floor.

hugh

--------------------
Technical Editor, Sound On Sound

Quote Hugh Robjohns:

---

Depends if you are talking about what the human ear/brain perceives, or what a broadband audio test set would measure.

The dither signal is a random noise, where as the audio signal is, to a degree, recognisable and predictable. As such, the human ear/brain is able to ignore the random noise and determine with accuracy both the frequency and amplitude of the signal below the noise floor.

---

And if you want to take things a stage further then look up noise shaped dither. That's one way of increasing the apparent resolution to gain better performance from fewer bits which relies on the ear being more sensitive to some frequencies than others.

Cheers

James.

--------------------
JRP Music - Audio Mastering and Restoration.
http://www.jrpmusic.net

Edited by James Perrett (27/11/06 03:27 PM)

Quote Hugh Robjohns:

---

The problem with this approach for high quality audio systems is that it is extremely difficult to processes non-linear quantised signals together. Because the quantising intervals vary you can nop longer simply add the bits! So non-linear quantising is only used in applications where convenience is more important than quality!

---

Indeed - the audio is compressed purely to make it easy and cheap to transmit and switch across a telecom network. In practice, I believe that processing on the encoded audio (eg echo-cancellation, data rate compression [ADPCM, LDCELP yeuch!],...) is done by internally decoding back to 12+1 format. Which is quite trivial to do in recent decades.

But isn't a similar (albeit one-way) non-linear technique used in the top couple of bits in the dbx type-IV ADC process?

Quote James Perrett:

---

And if you want to take things a stage further then look up noise shaped dither. That's one way of increasing the apparent resolution to gain better performance from fewer bits which relies on the ear being more sensitive to some frequencies than others.

---

And this always seems just a little bit like employing the kind of dirty tricks much criticised in perceptual coding systems! Effective as it may be..

--------------------
TAKEN.BALL.GONE.HOME

Quote Hugh Robjohns:

---

... It is better to imagine a ruler calibrated only in meters. Adding extra bits to the wordlength is the equivalent of adding centimetre divisions to the ruler. And adding more bits is like adding millimetre divisions. The additional accuracy isn't all in the first metre, but spread out across the entire length of the original ruler. Measurements made at any point on the ruler can be performed with greater accuracy (or less error). That translates in an audio converter to meaning that there is less noise, or a lower noise floor...

---

Thanks for this explanation Hugh, couldn't quite understand the extra bits business until you explained it like this.

Quote JimiQ:

---

But isn't a similar (albeit one-way) non-linear technique used in the top couple of bits in the dbx type-IV ADC process?

---

As far as I know, this is done to the analogue signal before it gets anywhere near the converter. It has to be done in the analogue domain, otherwise you would need a ridiculous amount of headroom in the converter!

Quote:

---

And this always seems just a little bit like employing the kind of dirty tricks much criticised in perceptual coding systems! Effective as it may be..

---

Not really. Perceptual coders are throwing away audio signal they think you won't be able to hear.

Noise shaping is exactly that -- altering the spectral content of the dither noise signal to minimise it's audibility, generally by reducing the level in the 2-6kHz region. It is the broadband average level that is critical for dither to work, so scopping out some energy in the areas where we are most sensitive, and building it up in areas where we are less sensitive makes a significant subjective improvement.

I have some examples of a dynamic piano piece recorded with 3 bit quantising without dither, with spectrally flat dither, and with noise-shaped dither. The first is unlistenable. The second is drowned completely in noise, and the third sounds like a low level cassette recording -- hissy but listenable!

hugh

--------------------
Technical Editor, Sound On Sound

I've struggled with this stuff for a loooong time now - every now and again I get a major mental breakthrough then find-out that I've got it wrong. Again!

I have always struggled profoundly with the idea that 24 bit doesn't translate to smoother transitions between different volumes than 16 bit. I've slightly grasped dither as being something that is used when changing word length, but I'm currently thinking that it sounds as though ALL digital signals are dithered so as to render a linear apparent volume response, and that because 24 bit wordlengths produce smaller steps they need less dither = less noise.

Go one, someone ruin my weekend. Again!!

(do love this forum though, despite the head pains!!!)

Andi

Edited by Andip (07/09/07 03:45 PM)

Thats pretty much right, but note that real world 24 bit signals don't need dither in general as the noise floor of pretty much any real world microphone/preamp/ADC chain will bury the lowest few bits in hiss even without dither.

Regards, Dan.

--------------------
Audiophiles use phono leads because they are unbalanced people!