Dirac upcoming technologies for non linearities

[Davide]

I came across this recent Dirac blog page.

They talk about extremely interesting things, such as technology called "Non Linear Control".

Here is an excerpt:

Our nonlinear control (NLC) technology, which we will present at AES 2024, uses machine learning to enhance loudspeaker performance while reducing the need for heavy and expensive components. By employing Long Short-Term Memory (LSTM) neural networks, we can make the driving force on the voice coil (the part of the speaker that turns electricity into sound) more consistent, improving the mechanical design and compensating for magnetic limitations.

NLC adjusts the voice coil current to correct force factor irregularities (inconsistencies in the voice coil’s efficiency) without requiring complex mechanical measurements. In tests with an otherwise suboptimal driver, our technology reduced distortion by 10 dB, nearly matching the performance of a well-designed driver.

Our NLC technology simplifies audio systems by eliminating the need for feedback loops, sensors, and feedback mechanisms, which cuts costs and enhances sustainability. This makes high-quality audio systems more affordable and environmentally friendly.

A key factor of our NLC technology is its real-world effectiveness, where factors like temperature and aging can impact speaker performance. The LSTM neural networks get ahead of these changes at manufacture time, ensuring consistent audio quality in a range of external conditions. This robustness makes NLC ideal for automotive and portable audio applications, where environmental conditions vary widely.

Additionally, by focusing on voice coil current and force factor nonlinearity, our NLC technology targets the most stable and predictable aspects of speaker operation. This approach eliminates the need for complex mechanical signal acquisition, such as displacement or velocity measurements, which are costly and difficult to implement in production.

On a technical level it seems extremely interesting to me.

What do you think?

To me is not clear if it is a domestic target.

@Flak Can you add something?

 

[neRok]

Sounds like a good idea if it can be reduced to a simple algorithm.

 

[Davide]

Sounds like a good idea if it can be reduced to a simple algorithm.

Yeah, it is partly what Purifi has aimed to achieve via hardware with its drivers. If I don't misinterpret.

 

[Holmz]

I came across this recent Dirac blog page.

They talk about extremely interesting things, such as technology called "Non Linear Control".

Here is an excerpt:


On a technical level it seems extremely interesting to me.

What do you think?

To me is not clear if it is a domestic target.

@Flak Can you add something?

Sounds like a feed forward model.
Chucking in “Nueral network” is like saying AI and then gain marketing points..
But one could compute/calibrate it.

Yeah, it is partly what Purifi has aimed to achieve via hardware with its drivers. If I don't misinterpret.

If you look at the BL curve, it is move linear for a purify drver… so the force on the cone is more linear
So the non linear distortions become lower.

If one knows the curve is not linear, then one could adjust the current to make the force linear, and get rid of non linear distortions.

And one can do both the more linear driver, and then correcting its non-linearity..

It should help the most at higher SPL, as that is where the VC is sneaking out of the feild. And it should help the most with “Long throw” drivers, like subwoofers, and reduce their harmonics.

 

[Davide]

Sounds like a feed forward model.
Chucking in “Nueral network” is like saying AI and then gain marketing points..
But one could compute/calibrate it.


If you look at the BL curve, it is move linear for a purify drver… so the force on the cone is more linear
So the non linear distortions become lower.

If one knows the curve is not linear, then one could adjust the current to make the force linear, and get rid of non linear distortions.

And one can do both the more linear driver, and then correcting its non-linearity..

It should help the most at higher SPL, as that is where the VC is sneaking out of the feild. And it should help the most with “Long throw” drivers, like subwoofers, and reduce their harmonics.

Two things are curious in my opinion:

"...without requiring complex mechanical measurements."

It would be interesting to know if and how the neural network is able to determine non-linearities without any initial measurement.

And:

"A key factor of our NLC technology is its real-world effectiveness, where factors like temperature and aging can impact speaker performance. The LSTM neural networks get ahead of these changes at manufacture time, ensuring consistent audio quality in a range of external conditions."

I understand that the Neural Network can well predict how the suspension will age. And it can create a very valid model of how non-linearities vary according to environmental variables such as temperature.
But precisely in the case of temperature, for example, a sensor should be required at least for ambient temperature to give an input to the model and get the appropriate mathematical function to correct the non-linearities of that moment.

Of course, nothing additional in automotive ...

 

[Holmz]

I dunno…
A friend and I were talking about this topic a couple of years ago, and while we were somewhat keen, Mike decided skiing was a better use of time, and I opted to straddle the bicycle.

It seems pretty tractable to use some multi tones, and just determine what BL curve would produce the observed harmonics and IMD.
Then use those coefficients to determine the delta between the force that the cone is really exerting, and the force it would extent in ideally.
And supply the modified voltage.

It is a close a parallel to using S/W to correct lens distortions.
Yeah one could get a more expensive lens, or get a lens that was good in some things like having low chromatic abborations, and then taking out barrel distortion and vignetting with the S/W. So it can make cheaper lens perform better, but doesn’t do a lot for a great lens. (I.e. Nothing left to do.)

Two things are curious in my opinion:

"...without requiring complex mechanical measurements."

It would be interesting to know if and how the neural network is able to determine non-linearities without any initial measurement.

Maybe a microphone that samples the output?

 

[kemmler3D]

Sounds really plausible to me, if you sample the output of tones or multitones and look at the harmonic distortion, then basically do a lookup table / waveshaper to counteract that (maybe using an iterative machine learning approach)... why not? Or maybe some other approach entirely, but I don't doubt they could do this.

What I really want to know is whether it has any benefit for drivers that are already pretty good, or whether it's only useful for correcting high distortion. If you could get another 10dB out of (say) a Purifi 10", you're verging on inaudible distortion across the board out of an actual physical speaker. That's the dream...

 

[dlaloum]

Sounds really plausible to me, if you sample the output of tones or multitones and look at the harmonic distortion, then basically do a lookup table / waveshaper to counteract that (maybe using an iterative machine learning approach)... why not? Or maybe some other approach entirely, but I don't doubt they could do this.

What I really want to know is whether it has any benefit for drivers that are already pretty good, or whether it's only useful for correcting high distortion. If you could get another 10dB out of (say) a Purifi 10", you're verging on inaudible distortion across the board out of an actual physical speaker. That's the dream...

Could this potentially work alongside something like Dirac Live? An additional set of measurements at calibration time, and an additional layer of distortion reducing software....