6
$\begingroup$

Note: this question has been edited to refer to a sample that does not use dithering.

I recently asked my first question on here, about what I saw in an uncompressed audio file written out as integers. I learned about dithering.

This time, rather than silence, I'm examining a sine wave I generated. In Audacity, I generated a 700 Hz sine wave for 1 second, at amplitude 0.4. The bit rate is only 8000 samples per second.

I thought that when converted the 16-bit signed PCM mono WAVE audio to a list of 16-bit signed integers, that I would see a list of integers corresponding to the values I thought I'd see based on a simple calculation.

Here are the first 80 samples. Upon manual inspection, I found there to be a period of 80 samples, so these 80 samples just repeat:

 1.      0       41.      0 
 2.   6592       42.  -6848 
 3.  11423       43. -11679 
 4.  13067       44. -13323 
 5.  10604       45. -10860 
 6.   4760       46.  -5016 
 7.  -2306       47.   2050 
 8.  -8768       48.   8512 
 9. -12466       49.  12210 
10. -12745       50.  12489 
11.  -9524       51.   9268 
12.  -3060       52.   2804 
13.   3794       53.  -4050 
14.   9711       54.  -9967 
15.  12690       55. -12946 
16.  12109       56. -12365 
17.   7704       57.  -7960 
18.   1028       58.  -1284 
19.  -6207       59.   5951 
20. -11176       60.  10920 
21. -13363       61.  13107 
22. -11176       62.  10920 
23.  -6207       63.   5951 
24.   1028       64.  -1284 
25.   7704       65.  -7960 
26.  12109       66. -12365 
27.  12690       67. -12946 
28.   9711       68.  -9967 
29.   3794       69.  -4050 
30.  -3060       70.   2804 
31.  -9524       71.   9268 
32. -12745       72.  12489 
33. -12466       73.  12210 
34.  -8768       74.   8512 
35.  -2306       75.   2050 
36.   4760       76.  -5016 
37.  10604       77. -10860 
38.  13067       78. -13323 
39.  11423       79. -11679 
40.   6592       80.  -6848

Based on the Nyquist frequency (4000 Hz), I calculated the values that I expected to see when I opened the file:

$v(n)=aM \sin\left({\dfrac{n\pi f_\mathrm{data}}{f_\mathrm{Nyquist}}}\right)$

$v(n)$ is the value of the sample written out as a 16-bit signed integer

$a$ is the amplitude of the sine wave (here, it is 0.4)

$M$ is the maximum possible value written out as a 16-bit signed integer (here, it is 32,767)

$n$ is the index of the sample (the first sample has an index of 0, the next is 1, etc.)

$f_\mathrm{data}$ is the frequency of the sine wave (here, it is 700)

$f_\mathrm{Nyquist}$ is the Nyquist frequency of the signal (here, it is 4000)

According to the above formula, I thought I would see these as the first 12 samples instead. Note: due to rounding/truncating, the program I used to tabulate these values differs from the formula above by [-1, +1].

 1.      0       41.      0 
 2.   6848       42.  -6848 
 3.  11678       43. -11678 
 4.  13066       44. -13066 
 5.  10603       45. -10603 
 6.   5015       46.  -5015 
 7.  -2050       47.   2050 
 8.  -8512       48.   8512 
 9. -12465       49.  12465 
10. -12744       50.  12744 
11.  -9267       51.   9267 
12.  -3059       52.   3059 
13.   4050       53.  -4050 
14.   9966       54.  -9966 
15.  12945       55. -12945 
16.  12109       56. -12109 
17.   7703       57.  -7703 
18.   1028       58.  -1028 
19.  -5950       59.   5950 
20. -11175       60.  11175 
21. -13106       61.  13106 
22. -11175       62.  11175 
23.  -5950       63.   5950 
24.   1028       64.  -1028 
25.   7703       65.  -7703 
26.  12109       66. -12109 
27.  12945       67. -12945 
28.   9966       68.  -9966 
29.   4050       69.  -4050 
30.  -3059       70.   3059 
31.  -9267       71.   9267 
32. -12744       72.  12744 
33. -12465       73.  12465 
34.  -8512       74.   8512 
35.  -2050       75.   2050 
36.   5015       76.  -5015 
37.  10603       77. -10603 
38.  13066       78. -13066 
39.  11678       79. -11678 
40.   6848       80.  -6848

The difference is much more than the [-7, +7] that I found in the scenario mentioned in my previous question.

In examining the differences between the values, I found this:

 1.      0       41.      0 
 2.    256       42.      0 
 3.    255       43.      1 
 4.     -1       44.    257 
 5.     -1       45.    257 
 6.    255       46.      1 
 7.    256       47.      0 
 8.    256       48.      0 
 9.      1       49.    255 
10.      1       50.    255 
11.    257       51.     -1 
12.      1       52.    255 
13.    256       53.      0 
14.    255       54.      1 
15.    255       55.      1 
16.      0       56.    256 
17.     -1       57.    257 
18.      0       58.    256 
19.    257       59.     -1 
20.      1       60.    255 
21.    257       61.     -1 
22.      1       62.    255 
23.    257       63.     -1 
24.      0       64.    256 
25.     -1       65.    257 
26.      0       66.    256 
27.    255       67.      1 
28.    255       68.      1 
29.    256       69.      0 
30.      1       70.    255 
31.    257       71.     -1 
32.      1       72.    255 
33.      1       73.    255 
34.    256       74.      0 
35.    256       75.      0 
36.    255       76.      1 
37.     -1       77.    257 
38.     -1       78.    257 
39.    255       79.      1 
40.    256       80.      0

A negative number indicates that the actual value is lower than I expected, and vice versa.

Here is the difference as a graph:

Raw data graph

It's kind of hard to read due to the scale, so here is a provisional graph where [255, 257] is mapped to [-1, 1]:

Graph that ignores the 256 offset

And here is a graph denoting which samples are offset:

Graph that only observes the 256 offset

Surely, this can't be dithering if I turned it off.

Again, the [-1, +1] is probably rounding/truncating in going from a floating-point type to an integer type. So what is the 256 there for?

More saliently, if I were to generate a sine wave from a list of integers ab initio, but didn't do this weird 256 stuff, would there be any issues due to those not being there?

For context, I can convert between 16-bit signed PCM WAVE audio and 16-bit signed integers bidirectionally. I'm using uncompressed audio right now to work out these kinds of issues. My end goal is some kind of embedded systems application, hence the desire to examine things at a numerical level. Before I switch to compressed audio, I want to ensure that every possible feature works with uncompressed audio because I can easily read it.

$\endgroup$
12
  • $\begingroup$ Sanity check first: if you disable dithering in the settings, do the value make sense? $\endgroup$ Feb 27, 2023 at 10:07
  • $\begingroup$ The problem is you assume how Audacity works when it generates or processes some piece of data you ask it to, but it may use completely different algorithm for generating the sine wave than what you assume, or process it using a completely different algorithm you expect it to. If you go to Audacity source code, you should be able to see how exactly it generates tones or handles the audio. $\endgroup$
    – Justme
    Feb 27, 2023 at 10:40
  • $\begingroup$ Can you add a plot the error difference you see over a much longer time duration? $\endgroup$ Feb 27, 2023 at 12:07
  • $\begingroup$ @MarcusMüller I'm about to edit my question regarding a dithering-disabled file comparison. $\endgroup$ Feb 27, 2023 at 21:06
  • 1
    $\begingroup$ @PatrickO'Brien: Still can't reproduce. The wave file I get out of Audacity is perfectly fine and doesn't match what you post. So I'm guessing there is a problem with your wave import (looks like a bit error somewhere). Can you post the wave file and your import code ? $\endgroup$
    – Hilmar
    Feb 28, 2023 at 1:36

2 Answers 2

6
$\begingroup$

Sorry, can't reproduce. I did the same thing with the dither turned off and found a perfect match between the generated sine wave and the calculated values. Even with a dither turned on things were very close (within +- 4).

If you want to dive into the theory behind Audio Signal Processing, I suggest using a tool like Matlab or Python which are specifically designed for scientific computing and for being mathematically precise. Audacity is a Digital Audio Workstation primarily for music production and editing. It's designed around "what sounds best" and not all details behind this are fully documented or guaranteed to be immutable.

$\endgroup$
2
  • $\begingroup$ So, does this mean that if I don't care about "what sounds best" then I can generate my sine waves with just the mathematical function and convert it to audio all the same? $\endgroup$ Feb 27, 2023 at 22:54
  • 1
    $\begingroup$ No. What it means is "don't expect precise mathematics from Audacity". If you want to do science, use Matlab or Python. In most cases the results will be similar or exactly the same, but Audacity doesn't guarantee that (or doesn't provide full documentation of the entire process) $\endgroup$
    – Hilmar
    Feb 27, 2023 at 23:48
1
$\begingroup$

Given the offset is exactly 256+/-1 and not occurring on every value suggests the 8th bit in from the LSB is stuck, always 0 or always 1 on one of the waveforms so that the resulting error is 256. Review the binary value specifically for both waveforms at each time increment to confirm if this is the case.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.