Great video

The "null test" at around 21:40 is of particular interest to me although I can imagine a few audiophools screaming obscenities at their monitors  

Before AI came along, we used to remove the vocals from songs to create backing tracks. As the vocals are usually positioned right in the centre of the stereo recording, they are equally loud on both sides. We’d invert one of the stereo channels and then mix it into the other channel.

  Martin H. said  

Before AI came along, we used to remove the vocals from songs to create backing tracks. As the vocals are usually positioned right in the centre of the stereo recording, they are equally loud on both sides. We’d invert one of the stereo channels and then mix it into the other channel.

Sure, there are lots of apps like this one and they are quite amazing. Personally, I am not satisfied with the results and so I load them into a DAW (digital audio workstation) as separate tracks and replace them one at a time.

The interesting thing about the video, for me, is the fact that they demonstrate that a higher sampling rate doesn't mean "better".

I have the same beef with servo-motion-control; many years ago, it was established that a 1KHz PID sample-time was the sweet-spot and of course, AI agrees, only because it has become a belief. My real-world experiments don't agree. What puzzled me is that;

  Quote  

For example, audio CDs have a sampling rate of 44100 samples/second. At 0.5 cycle/sample, the corresponding Nyquist frequency is 22050 cycles/second (Hz). Conversely, the Nyquist rate for sampling a 22050 Hz signal is 44100

However, for a closed-loop servomotor, they recommend five to ten times the bandwidth of the motor. Not the Nyquist two-times.

The definition of bandwidth is the time taken to reach ~63% of maximum velocity.

So if a 3000RPM motor can achieve 1800RPM in 0.01 seconds, it has a bandwidth of 100Hz. Therefore, "ten times the bandwidth" means a 1KHz PID sample rate.

In my world, a small motor is 1HP and it is mechanically coupled via belts or ball-screws or gear reducers to a load that has inertia/friction/stiction, etc. (mechanical LPF). So in my case and the majority of cases, the numbers are overkill. They waste processing power and I have found that lowering the sample-rate to 500Hz and 250Hz loses no performance and seems to result is easier loop tuning (only a feeling at this point).

Over the years, I have read countless publications pertaining to PID motor control but I am starting to suspect that many of these "authors" have never had actual experience with a real motor and are merely parroting/plagiarizing.

PID is a mathmatical calculation using proportional x, integral x, and differential x.
As such it is essentially a relation of third order. You can compare that to a electronic filter (i.e. LCL).

In a filter that results in a maximum roll off of 18dB per octave. Such is the picomite audio filter.

But real CD player filters that output 20kHz at 44.1kHz samplerate, require more than 100dB per octave, so they are not of order 3, but more like order 15. Either hardware, or multitap FIR

Similar you could work with a lower PID bandwidth if you created a PIIIIDDDD

Disadvantage is that the lower bandwidt, and increased number of taps, will slow down the step response. In analogy with the LCL filters: the group-delay will increase.

Volhout
Edited 2026-05-31 04:13 by Volhout