• Register
  • Help
Page 1 of 2 12 LastLast
Results 1 to 10 of 20

Topic: Dumb audio question of the day (loss through normalization?)

Share/Bookmark
  1. #1

    Dumb audio question of the day (loss through normalization?)

    Let's say I have a tune and its peak volume is -0.5dB. If I normalize it down to, say, -3dB, and then back up to -0.5dB later, there's no loss of any kind, correct? (assuming I'm working on a raw WAV). Of course if I normalize to something really low like -30dB, and then raise it back up, I might've lost some softer parts of the piece. But I'm talking just a matter of 6-10dB or so.

    You might wonder why I don't already know the answer to something this simple. I used to know, but aliens stole it from my brain.

    Thanks
    Sam Hulick
    Composer
    http://www.samhulick.com/

  2. #2

    Re: Dumb audio question of the day (loss through normalization?)

    Forgive me but my forte is not physics or acoustics, but I'd have to say that anytime you normalize down you will be losing information/quality somewhat. Granted, if you take a 24 bit file and normalize down from -0.5 to -3 that quality loss will probably be very minimal but it is a loss. Take for instance the extreme. You normalize from -0.5 down to -50 or -70 or -infinity. You then have a very real quality loss. Normalizing to -3 is just a much less significant quality loss.

    Please correct me if I'm wrong.

    I think the real question is if this is an acceptable quality loss. I'd say I probaby wouldn't have a problem doing it to my music if necessary, especially if it were 24 bit or greater wav files.

    fizbin

  3. #3

    Re: Dumb audio question of the day (loss through normalization?)

    Hey Nick, I don´t mind real or PDF, but if I have to pay a more I don't mind, because I think it worths.
    Eduardo Tarilonte |Sample Library Developer
    SampleLibraries |MyWeb

  4. #4

    Re: Dumb audio question of the day (loss through normalization?)

    Every 6 decibels will lose you a bit. Changing the gain first by -48 dB, saving the file as 24 bit audio and then changing back +48 dB makes your 24 bit audio truncated 16 bit audio with 8 padded zeros. Changing the gain first by -96 dB and then +96dB makes your 24 bit audio truncated 8 bit audio with 16 padded zeros. Changing the gain first by -144 dB and then +144 dB makes your 24 bit audio 0 bit audio (silence) with 24 padded zeros.

    If the internal calculations of the audio processor are done in 48 bits the processor has an internal dynamic range of 288 decibels. If you change the gain by -144 dB and then +144 dB WITHIN such a processor the result will be identical to the original audio as long as the source audio has 24 or less bits.

    You won't notice the difference when the changes are small enough but the audio gets slightly degraded. You should also clarify the meaning of dithering to yourself. Bob Katz makes a brilliant explanation in his Mastering Audio. These things are meant to be the last phases of your production, don't do unnecessary gain changes in the middle of the project.

  5. #5

    Re: Dumb audio question of the day (loss through normalization?)

    Quote Originally Posted by MDesigner
    Let's say I have a tune and its peak volume is -0.5dB. If I normalize it down to, say, -3dB, and then back up to -0.5dB later, there's no loss of any kind, correct? (assuming I'm working on a raw WAV). Of course if I normalize to something really low like -30dB, and then raise it back up, I might've lost some softer parts of the piece. But I'm talking just a matter of 6-10dB or so.
    In theory, every normalization process will degrade the quality somewhat. Consider this: the signal paths of most modern DAWs process audio data in floating point format, usually with 64bit precision. When rendering an audio file, the last step of the signal flow will be a wordlength reduction of this data to, say, 16bit integer (usual WAV format). Prior to this step, the signal will (hopefully) be dithered, which basically means that random noise (well, not so random in reality, but bear with me) will be added around the level of the output format's Least Significant Bits (at 16 bit, this level is around -90dB) in order to mask the quantization noise, which is a byproduct of this wordlength reduction.

    When you change the level of the resulting file after this process, two things happen:

    - The wordlength of the signal will be increased again. Mathematically spoken, a gain change is a multiplication, so the result of this operation will most likely exceed the precision range of the source wordlength. An example: say you want to lower the level of a 16bit audio signal by 3dB. Your software will accomplish this by multiplying every sample value by a factor of about 0.7. Very soon, it'll stumble upon fractional results like 1828.4 - which is a problem, because this value can't be expressed in a 16bit integer format without losing the fractional part! So it needs to be rounded or truncated first (provided you want to keep the result at 16bit, that is), which will result in additional quantization distortion. You can somewhat diminish this problem by ensuring that your gain function will properly re-dither the resulting signal before reducing the wordlength again.

    - If the source signal was dithered before being quantized, the dither noise will be affected by the gain change aswell. This is usually not a problem when attenuating, since the original quantization noise will mostly be lost and replaced by the new one (see above), but when increasing the gain, you'll also raise this noise level. This might be only be critical for high gain values, though.

    Well, this is the theory at least. Of course, the interesting question here is "can the degradation be heard". I'd say this depends mostly on your source material. With 16bit, the degradation could probably be audible, especially at lower levels. With 24bit, probably not. In any case, you should try to keep the subsequent processing steps on "finished" audio files to a minimum and ensure proper re-dithering.

    I hope this helps somewhat...

    Cheers,
    jan

  6. #6

    Re: Dumb audio question of the day (loss through normalization?)

    Hi,

    Quote Originally Posted by janila
    Every 6 decibels will lose you a bit. Changing the gain first by -48 dB, saving the file as 24 bit audio and then changing back +48 dB makes your 24 bit audio truncated 16 bit audio with 8 padded zeros. Changing the gain first by -96 dB and then +96dB makes your 24 bit audio truncated 8 bit audio with 16 padded zeros. Changing the gain first by -144 dB and then +144 dB makes your 24 bit audio 0 bit audio (silence) with 24 padded zeros.
    Sorry to be nitpicking - in a properly dithered system, those numbers are actually a bit different. Try this in a DAW: you can actually attenuate a signal (peak at 0dBfs) by -96dB, output it to a 16bit format, re-import it, boost it by 96dB and still make out the original signal (how well depends on the quality of the dithering algorithm). The result isn't exactly pretty, but quite eye-opening (or ear-opening?), as one can hear the characteristics of different dithering/noise shaping algorithms quite well this way.

    And yeah, I know the numbers are close enough to get the point across, I'll crawl back under my stone now

    Cheers,
    jan

  7. #7

    Re: Dumb audio question of the day (loss through normalization?)

    Quote Originally Posted by Jan Morgenstern
    Sorry to be nitpicking - in a properly dithered system, those numbers are actually a bit different. Try this in a DAW: you can actually attenuate a signal (peak at 0dBfs) by -96dB, output it to a 16bit format, re-import it, boost it by 96dB and still make out the original signal (how well depends on the quality of the dithering algorithm). The result isn't exactly pretty, but quite eye-opening (or ear-opening?), as one can hear the characteristics of different dithering/noise shaping algorithms quite well this way.
    I was referring to gain change without dither as was the original poster. Without dither the numbers are exact. Dithering must be done only once at the end of the chain and MDesigner was asking about a situation where only the gain change is affecting the signal. That's why I said that he has to understand the meaning of dithering to know what the real limitations are. I was just trying to make the thing simple.

  8. #8
    Senior Member Bruce A. Richardson's Avatar
    Join Date
    Sep 1999
    Location
    Dallas, Texas
    Posts
    5,755

    Re: Dumb audio question of the day (loss through normalization?)

    Quote Originally Posted by MDesigner
    Let's say I have a tune and its peak volume is -0.5dB. If I normalize it down to, say, -3dB, and then back up to -0.5dB later, there's no loss of any kind, correct? (assuming I'm working on a raw WAV). Of course if I normalize to something really low like -30dB, and then raise it back up, I might've lost some softer parts of the piece. But I'm talking just a matter of 6-10dB or so.

    You might wonder why I don't already know the answer to something this simple. I used to know, but aliens stole it from my brain.

    Thanks
    The larger question is why you would do this. At issue is not so much the normalization process itself, but the quantization steps which occur every time you run a "destructive" process.

    If you could elaborate on the situation you're facing, that might yield some responses which could help suggest other methodologies.

    At first blush, if you are tweaking volumes of mixes around in this manner, I would assume some larger structure that you are plugging these into. In that case, I would suggest leaving the original file unaltered and doing this "leveling" work in a non-destructive editor, spitting out new files at the different levels. That way, your original is always there--its purity remains no matter how many different levels you've subsequently produced.

  9. #9

    Re: Dumb audio question of the day (loss through normalization?)

    Hi Janila,

    Quote Originally Posted by janila
    Dithering must be done only once at the end of the chain and MDesigner was asking about a situation where only the gain change is affecting the signal.
    but that's been my point - to be effective, dithering has to be applied at _every_ point in the chain where the wordlength has to be reduced, and changing gain in a non-floating-point environment (such as editing destructively within a sample editor) is one of those cases. Provided the processing takes place inside a floating-point DAW (that is, within the DAW's signal chain, not its sample editor ), you're perfectly right, though.

    Cheers,
    jan

  10. #10

    Re: Dumb audio question of the day (loss through normalization?)

    Quote Originally Posted by janila
    Dithering must be done only once at the end of the chain and MDesigner was asking about a situation where only the gain change is affecting the signal.
    Quote Originally Posted by Jan Morgenstern
    but that's been my point - to be effective, dithering has to be applied at _every_ point in the chain where the wordlength has to be reduced
    Yes I have to rephrase my statement. Wordlength must be shortened only once, at the end of the signal chain. Dithering must be done only once at the end of the chain while shortening the wordlenght (if the wordlenght needs to be shortened). The devices that use higher internal bit depths are going to take care of their dithering/rounding and those cases don't affect the signal significantly as the errors are atleast 138 decibels below the digital zero.

    changing gain in a non-floating-point environment (such as editing destructively within a sample editor) is one of those cases.
    Yes and no. I can't figure out a situation where the gain would need to be lowered like that. I would rather say that it shouldn't be done at all as Bruce suggested above.

Go Back to forum

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •