Converting 16bit 44.1khz audio to 8bit 44.1khz, Correct method?

[Truebelieff]

Hi all, Its Time to nerd out :

So I have a 16bit 44.1khz wav, with intention to convert it to 8bit 44.1khz w/ least SQ reduction possible.
The method I thought of is:
An attempt to demonstrate the method in chart:
[16bit, 44.1Khz] -----> {Upsample w/ SOX to 384KHZ} -----> [16bit, 384Khz] -----> {Reduce Bitdepth w/ R8Brain w/ Strong Noise Shaping to 8bit} -----> [8bit, 384khz, Dithering quantization noise mostly residing in the 100th of khz] -----> {Downsample w/ SOX TO 44.1KHZ} -----> [8bit, 44.1khz, Dithering quantization noise *which resided* in the 100th of khz before- IS NOW GONE] -----> #"MAGIC" REMOVAL OF MOST DITHERING_QUANTIZATION_ERROR_IN_SPREAD_TO_BACKGROUND_NOISE_FORM?#

Non chartedd 'dry' explanation of the method:
A] Convert the 16bit 44.1khz wav to very high sample rate in Reaper w/ SOX, say 16bit 384khz wav.
B] Convert the 16bit 384khz file to 8bit 384khz, Using R8Brain Dithering w/ stongest noise shaping possible to move the dither noise completely out of audible frequancies.
C] Then downsample the 8bit 384khz wav to 8bit 44.1khz using SOX again.
D] Result: 8bit 44.1khz wav w/ theleast audible dithering and quantization noise possible (?)
Will the 3rd step eliminate dither noise while retainig the ultra low quantization error left from R8Brain convertion at the 2nd step?

Am I mistaken in my understanding of matigiating the loss of SQ when converting the 16bit 44.1khz wav to 8bit 44.1khz wav using the 3 steps I've described?

 

[SIY]

Hi all, Its Time to nerd out :

So I have a 16bit 44.1khz wav, with intention to convert it to 8bit 44.1khz w/ least SQ reduction possible.
The method I thought of is:
A] Convert the 16bit 44.1khz wav to very high sample rate in Reaper w/ SOX, say 16bit 384khz wav.
B] Convert the 16bit 384khz file to 8bit 384khz, Using R8Brain Dithering w/ stongest noise shaping possible to move the dither noise completely out of audible frequancies.
C] Then downsample the 8bit 384khz wav to 8bit 44.1khz using SOX again.
D] Result: 8bit 44.1khz wav w/ theleast audible dithering and quantization noise possible (?)

Will the 3rd step eliminate dither noise while retainig the ultra low quantization error left from R8Brain convertion at the 2nd step?
Am I mistaken in my understanding of matigiating the loss of SQ when converting the 16bit 44.1khz wav to 8bit 44.1khz wav using the 3 steps I've described?

Why are you trying to do this? It creates absolutely no new information and loses a pile of dynamic range.

 

[Truebelieff]

@SIY: I'm trying to eliminate as far as possible the added dithering information... while keeping the benefits of dithering's spreading the quantization error equally as background noise. So eliminating quantization noise is basically free removal of quantization error?

@Blumlein 88:8 bit files (which aren't converted from loudness wars's victimized 16bit wavs) are generally very noisy in their background when dithering. I'm thinking to remove that noise as much as possible while keeping the quantization error at the lowest value.
Doesn't my method provide any audible benefits compared to direct converting to 8bit w/ sox w/ dithering as suggested by Blumlein 88?

 

[Blumlein 88]

I think SIY's question is why convert to 8 bit? Are you trying to save file size or what do you have in mind? Maybe if you tell us that there is a better way to accomplish what you want without going to 8 bit.

 

[Truebelieff]

8bit is per request from me (I guess for use in custom modded game for some old console)... I know it suboptimal bit depth but higher bit depth isn't an option. So I'm trying to make the best of it... Am I?

 

[DVDdoug]

Increasing the sample rate won't help. Quantization noise is related to bit depth...

Dither is added noise that's supposed to sound better than quantization noise alone. You can try with and without dither and you can try different dither algorithms.

Quantization noise goes-away completely with digital silence whereas dither does not (unless you can find a "smart dither" algorithm that doesn't dither silence).

 

[Truebelieff]

Increasing the sample rate won't help. Quantization noise is related to bit depth...

Dither is added noise that's supposed to sound better than quantization noise alone. You can try with and without dither and you can try different dither algorithms.

Quantization noise goes-away completely with digital silence whereas dither does not (unless you can find a "smart dither" algorithm that doesn't dither silence).

And what if I use strong noise shaping to move most of the noise to high sample rates in the hundred of khz when reducing to 8bit depth, and then downsampling that file to to 44.1khz while most of the dithering noise left in the hundreds of khz is eliminated?

 

[DVDdoug]

8 bit files (which aren't converted from loudness wars's victimized 16bit wavs) are generally very noisy in their background when dithering...

... 8bit is per request from me (I guess for use in custom modded game for some old console)

Use compression and limiting to keep the sound constantly loud to drown-out the quantization noise. Then don't dither and it will be silent during silence.

P.S.
And put the game in a noisy arcade!

 

[Blumlein 88]

I assume your thinking is shaped dither lowers noise in the audible band and pushes noise up into the higher frequencies. So by using shaped dither to 384 khz the highest dither noise will be well up above 20 khz. So then you can convert down to lower sample rates while having less noise in the file. I think you'll find even using shaped dither it will put the noise level at the same level when you do the conversion to 8 bit at the lower rate. Easy enough to try in any case. Record a low level tone and go thru both conversion processes in your software and see if their is any difference once finished.

PS just saw your reply confirming that is what you had in mind.

 

[Truebelieff]

@DVDdoug: One of mine objectives to keep the dithering noise as absent as possible w/o compression & w/o avoiding dither which prevents the much worser then anything Quantization Distortion.

@Blumlein 88 : You got my thinking absolutely right! But what if:
[24bit, 44.1Khz] -----> {Upsample w/ SOX to 384KHZ} -----> [24bit, 384Khz] -----> {Reduce Bitdepth w/ R8Brain w/ Strong Noise Shaping to 16bit} -----> [16bit, 384khz, Dithering quantization noise mostly residing in the 100th of khz] -----> {Downsample w/ SOX TO 44.1KHZ} -----> [16bit, 44.1khz, Dithering quantization noise *which resided* in the 100th of khz before- IS NOW GONE]
Will the upsampling be more beneficial when the dithering noise have additional 8bits to 'breath' compared to my Original Post?

 

[Blumlein 88]

@DVDdoug: One of mine objectives to keep the dithering noise as absent as possible w/o compression & w/o avoiding dither which prevents the much worser then anything Quantization Distortion.
@Bumlein 88: You got my thinking absolutely right! But what if:
[24bit, 44.1Khz] -----> {Upsample w/ SOX to 384KHZ} -----> [24bit, 384Khz] -----> {Reduce Bitdepth w/ R8Brain w/ Strong Noise Shaping to 16bit} -----> [16bit, 384khz, Dithering quantization noise mostly residing in the 100th of khz] -----> {Downsample w/ SOX TO 44.1KHZ} -----> [16bit, 44.1khz, Dithering quantization noise *which resided* in the 100th of khz before- IS NOW GONE]
Will the upsampling be more beneficial when the dithering noise have additional 8bits to 'breath' compared to my Original Post?

It is too simple to just do it and see what the result is. If you have the software handy for all of this you can do it both ways quicker than we can discuss it. Some processes like EQ and other DSP benefit from upsampling and downsampling the result. I don't think bit reduction and dither will, but try it and see. If I was in a place to do it this morning I'd do that myself.

 

[Hayabusa]

Increasing the sample rate won't help. Quantization noise is related to bit depth...

dither can shift the spectrum of the quantitation noise to the inaudible spectrum, so it does help

 

[Truebelieff]

Thanks all for the answers! I'll test double blindly for actual audible bebfits for my hypotesis sometimes in the future + Some intersting insight from Paid Gemini Pro AI on the matter:
"
Putting it all together

1. Dithering with Strong Noise Shaping at High Sampling Rate:
   • You start with a high-resolution audio file (e.g., 24-bit, 96 kHz).
   • You apply dithering along with strong noise shaping to reduce the bit depth (e.g., to 16-bit). The high sampling rate gives the noise shaping algorithm more room to work, pushing the noise energy into very high frequencies.

1. Downsampling:
   • You then reduce the sampling rate (e.g., to 44.1 kHz). The high-frequency noise introduced by the dithering process is now above the Nyquist frequency of the new, lower sampling rate.
   • During the downsampling process, this high-frequency noise is effectively filtered out, leaving you with a cleaner, lower-resolution audio file that retains much of the original quality.

Benefits

• Reduced File Size: Downsampling results in a smaller file size, which is useful for storage and streaming.
• Improved Audio Quality: Dithering and noise shaping minimize the audible artifacts associated with bit depth reduction, preserving the audio quality even after downsampling.

Potential Considerations

• Strong Noise Shaping Artifacts: In some cases, strong noise shaping can introduce subtle artifacts that might be audible in certain audio material or on high-end audio systems.
• Processing Overhead: This process requires additional computational resources compared to simple downsampling without dithering and noise shaping."

 

[DVDdoug]

During the downsampling process, this high-frequency noise is effectively filtered out, leaving you with a cleaner, lower-resolution audio file that retains much of the original quality.

Why add dither if you are going to filter it out? There will likely be some residual effect but it probably won't be optimum.

 

[Truebelieff]

Why add dither if you are going to filter it out? There will likely be some residual effect but it probably won't be optimum.

From what I understand, dithering noise isn't 'added' during dithring but is a mere *relatively equal* spreading of the quantization distortion 'energy' across the whole frequancy spectrum, hance make it quiter from the psychoacusticall human ear aspect. So filterting out that dithering noise means the best of both worlds: No quantization distortion and almloast non of it in spreaded, background noise, form.

 

[Hayabusa]

Why add dither if you are going to filter it out? There will likely be some residual effect but it probably won't be optimum.

Indeed if you end up with 44.1 your should dither to 44.1 else it wont be very effective

 

[Blumlein 88]

Thanks all for the answers! I'll test double blindly for actual audible bebfits for my hypotesis sometimes in the future + Some intersting insight from Paid Gemini Pro AI on the matter:
"
Putting it all together

1. Dithering with Strong Noise Shaping at High Sampling Rate:
   • You start with a high-resolution audio file (e.g., 24-bit, 96 kHz).
   • You apply dithering along with strong noise shaping to reduce the bit depth (e.g., to 16-bit). The high sampling rate gives the noise shaping algorithm more room to work, pushing the noise energy into very high frequencies.

1. Downsampling:
   • You then reduce the sampling rate (e.g., to 44.1 kHz). The high-frequency noise introduced by the dithering process is now above the Nyquist frequency of the new, lower sampling rate.
   • During the downsampling process, this high-frequency noise is effectively filtered out, leaving you with a cleaner, lower-resolution audio file that retains much of the original quality.

Benefits

• Reduced File Size: Downsampling results in a smaller file size, which is useful for storage and streaming.
• Improved Audio Quality: Dithering and noise shaping minimize the audible artifacts associated with bit depth reduction, preserving the audio quality even after downsampling.

Potential Considerations

• Strong Noise Shaping Artifacts: In some cases, strong noise shaping can introduce subtle artifacts that might be audible in certain audio material or on high-end audio systems.
• Processing Overhead: This process requires additional computational resources compared to simple downsampling without dithering and noise shaping."

You don't need a double blind test for this. Just do the files and do a spectrum analysis of the result. If you have nothing more than a low level tone, you also could notch it out after the fact and digitally amplify what is left to simply hear what the difference is.

 

[AnalogSteph]

Indeed. This plan is elaborate nonsense. Whatever happens, you need to have enough noise within a 22.05 kHz bandwidth at the end of the day, there's no way around it. You need one and only one step: dither 16/44 --> 8/44.

The perceptively best dither I've come across is Foobar2000's (which should run on WINE since I assume you're a Linux person, and it doesn't have to be the latest and greatest version; the converter and a playlist should make quick work of a bunch of files). I suggest you try both that and whatever shaped dither is available in SoX and listen which one you like best. (The one in Audacity is definitely an improvement over flat but it creates a noise peak around 10 kHz which is the area where hiss tends to be most noticeable.) 8 bits is low enough for things to be quite plainly audible, especially when starting from something quiet or silence.

 

[retroflex]

What you're trying to do fundamentally cannot work. What would the sample values be for a low bit depth signal with no quantization noise and no dither?

The quantization noise will come back when you filter out the high frequencies, because the sample values of the filtered signal cannot be accurately represented at the low bit depth.

 

[Hayabusa]

• Reduced File Size: Downsampling results in a smaller file size, which is useful for storage and streaming.

8 bits gives you a factor of 2.
You know FLAC will give almost the same with no loss?