Audacity version: 2.2.2
All I’m trying to do is to convert audio from this:
to this:
but however much I search online, there’s only results about people converting audio to PWM audio just for arduinos, but I want to be able to play back the PWM in audacity
Thanks,
8bit
For pulse width modulation to work, the switching frequency needs to be much higher than the highest waveform frequency that you want to represent. In the case of full-band audio, the switching frequency is often around 1 MHz. A normal sound card can’t handle such high frequencies.
If this is just an experiment, then it would be possible convert a very low frequency signal to PWM.
What’s the purpose of this?
Pretty much just an experiment. I have sample-rate of the audio track set to 0.5 MHz, so I just wanna mess around and experiment.
If you’ve ever programmed you might be able to do something with [u]Nyquist[/u].
I have sample-rate of the audio track set to 0.5 MHz,
Your screenshot says 48000Hz. 
I don’t know if you can play a file with such a sample rate. It’s a lot of data so you may not be able to read/process it fast enough, and it’s not standard so the drivers may not downsample it to something your soundcard can use.
If you can play it, PWM will probably work because it will be low-pass filtered.
The fake “PWM” that I had made by clipping some audio really loudly isn’t the sample rate i’ll be using, instead I can mess around with a PWM version of audio in slow playback speeds, I think it’ll sound interesting
A tracker plugin can produce pitch-tracking modulation of a square-wave …
Yes, you can do this in Nyquist (Nyquist - Audacity Manual), and it’s pretty simple to do 
In this example I’ll use a sample rate of 100000 Hz (100 kHz), just to make the numbers easier.
I’ll also use the “Nyquist Prompt” (Nyquist Prompt - Audacity Manual) for running the Nyquist commands.
I shall be using the current “v4” syntax, so the “Use legacy (version 3) syntax” checkbox must be empty (not ticked).
“PWM” is defined on Wikipedia (Pulse-width modulation - Wikipedia):
“Pulse-width modulation uses a rectangular pulse wave whose pulse width is modulated resulting in the variation of the average value of the waveform”
I shall use this definition, but not their described method of achieving it (because there’s a much easier way in Nyquist).
With a sample rate of 100 kHz, we can have a reasonably “square” looking waveform up to about 5 kHz. A 5000 Hz (5 kHz) waveform has a period of 1/5000th second. At a sample rate of 100000, that gives us (100000 * 1/5000) = 20 samples per cycle.
For each cycle of our 5 kHz pulse wave, we need to know what the average amplitude of our signal is. We can then generate a pulse wave where the pulse width is proportional to the average amplitude.
So now we need a test signal to work with.
I’ve chosen a “chirp” from 1 to 1000 Hz, amplitude 1.0 (start and end), Logarithmic interpolation, duration 5 seconds.
(snd-avg *track* 20 20 op-average)
track is the selected audio that is passed from Audacity to Nyquist
The first number is the number of samples that will be averaged in each step (the “block size”)
The second number is the number of samples that we move to get the next block of samples.
op-average tells the command to get the mean average of the samples in each block.
Thus the result will be a signal in which the first sample is equal to the average of the first 20 samples of the test tone, the second sample is the average of the next 20 samples, and so on.
We can assign the result to a variable using the SETF command:
(setf bias (snd-avg *track* 20 20 op-average))
We have set the value of “bias” to the result of (snd-avg track 20 20 op-average)
Now we can create our pulse wave using the Nyquist function OSC-PULSE
Remembering that our required frequency is 5000 Hz, the pulse generator command will be:
(osc-pulse 5000 bias)
where “bias” is the control signal that we created from SND-AVG.
So here’s the full code:
(setf bias (snd-avg *track* 20 20 op-average))
(osc-pulse 5000 bias)
and here’s the result:
Observations:
If you zoom in on the PWM waveform near the start of the track, you will see that the pulse wave is not always complete.
At the high frequency end of the “chirp”, the PWM signal only has about 5 pulses to define the waveform, so we can see that a 1000 Hz signal is approaching the limit of what a 5000 Hz pulse can handle.
Decoding:
If you try playing the PWM signal (turn the volume down a bit, it will be very loud), you will notice that it does not sound much like our original “chirp”. To recover our chirp sound, we need to decode the PWM signal. This is very easy to do. Just apply a low pass filter to remove the 5000 Hz carrier, and the result will be the “averaged” value of the pulses:
(lowpass8 *track* 1500)
LOWPASS8 is a steep lowpass filter.
The filter frequency of 1500 Hz is chosen to be far enough above our maximum signal frequency (1000 Hz) to not attenuate the signal too much, but well below our 5 kHz modulation frequency.
Note that at the low frequency end of the chirp, the demodulated output has distinct steps. This is because the 5000 Hz pulse wave has only 20 samples per cycle, so only 20 possible sizes for rectangular pulses, and because the change from one size rectangle to the next is so slow, our simple demodulator is unable to smooth out the steps.
Overall we can see that with a 100 kHz sample rate, and a 5 kHz pulse, our available audio bandwidth is around 200 Hz to 1000 Hz.
Wow, thanks! I was writing something in nyquist myself but it was giving me a completely wrong output, this makes much more sense!
Enjoy your experiments 
Hmmm, for some reason, putting the code in the nyquist prompt box results in this strange error output:
error: bad argument type - #(#<Sound: #5ace328> #<Sound: #5ace568>)
Function: #<Subr-SND-AVG: #aa4b810>
Arguments:
#(#<Sound: #5ace328> #<Sound: #5ace568>)
20
20
1
Function: #<FSubr-SETF: #aa46a30>
Arguments:
BIAS
(SND-AVG *TRACK* 20 20 OP-AVERAGE)
1> error: unbound variable - BIAS
if continued: try evaluating symbol again
Function: #<Subr-SND-AVG: #aa4b810>
Arguments:
#(#<Sound: #5ace328> #<Sound: #5ace568>)
20
20
1
Function: #<FSubr-SETF: #aa46a30>
Arguments:
BIAS
(SND-AVG *TRACK* 20 20 OP-AVERAGE)
2> 1>
From what I can tell, it doesn’t think track is a valid arguement, and it’s interpreting bias as a variable.
You’re using a stereo track (an array of two sounds).
The SND-AVG function requires a mono sound.
The easy workaround is to split the stereo track to mono (Splitting and Joining Stereo Tracks - Audacity Manual)
A better solution is to develop the code to handle stereo tracks. Some hints here: Missing features - Audacity Support
Ah, that makes more sense. Well, I’ve decided to make a nyquist plugin out of it and this is my code so far:
;nyquist plug-in
;version 3
;type process
;preview linear
;name "aPWM"
;action "Applying PWM"
;author "me"
;copyright "No copyright"
;control smp "Sample Bias" int "1-1000" 1 1 1000
;control htz "Carrier Frequency" int "1-20480000" 1 1 20480000
(setf bias (snd-avg *track* smp smp op-average))
(osc-pulse htz bias)
however, whatever I set in smp just ends up being the output, and debug doesn’t do any help either. I’m planning on making it automatically calculate a reasonable carrier frequency and sample bias.
It needs to be a “version 4” plug-in.
There were some important changes between versions 3 and 4. The biggest one being that in version 3 and earlier, the selected sound was passed from Audacity to Nyquist as the variable “S”. In version 4, the sound is passed from Audacity to Nyquist as the variable track.
Change the second line to:
;version 4