Music

Neural networks create authentic digital models of analog amps

Neural networks create authent...
In tests, listeners found digital emulations indistinguishable from the analog reference recording
In tests, listeners found digital emulations indistinguishable from the analog reference recording
View 1 Image
In tests, listeners found digital emulations indistinguishable from the analog reference recording
1/1
In tests, listeners found digital emulations indistinguishable from the analog reference recording

There's no denying that vintage guitar amps sound great, but they can be huge, expensive and prone to frequent breakdowns. Digital modeling has been getting better and better over the years, but there are folks who swear that the tones just don't sound authentic. Research out of Finland's Aalto University suggests otherwise.

There's nothing quite like the sound of an old tube amp to give your bottleneck blues playing an authentic vibe, or so the thinking goes. But profiling amps like the Kemper not only put a huge arsenal of vintage tones at your disposal, they sound pretty darn accurate too. So much so that they've attracted the attention of big hitters like Paul Gilbert, Pat Metheny and Steve Winwood.

Can such authenticity be reduced down to a plugin for music production software running on a laptop though? You could try to virtually model the complex circuitry of a boutique guitar amplifier, but this can be labor intensive and require a good deal of computer power to produce real-time results. Professor Vesa Välimäki, Alec Wright, Eero-Pekka Damskägg, and Lauri Juvela from the Aalto Acoustics Lab opted to make use of neural networks to lighten the load.

A digital model of the target amplifier was created by measuring the circuit's response to certain input signals and deep neural networks trained to emulate that amp. The researchers had previously modeled a Fender Bassman and three effects pedals using this method, but most recently trained two neural networks on the Blackstar HT-5 Metal and Mesa Boogie 5:50 Plus vacuum tube amps.

Then listening tests were performed, where 14 subjects – all of whom played at least one musical instrument – listened to short audio clips via the WebMUSHRA interface. The multiple stimuli with hidden reference and anchor (MUSHRA) method is often used to evaluate the perceived quality of audio compression algorithms.

The test subjects were played a reference clip followed by seven test clips from validation and test datasets through Sennheiser HD-650 headphones and asked to grade the accuracy of the timbre of the test tones against the reference tone.

The digital emulations for the Mesa Boogie amp were rated as "often indistinguishable from the reference" for both neural network samples. The results for the more distorted Blackstar emulations were not so clear cut. Ratings from one of the neural networks varied from good to excellent, while the other rated as excellent. And though all test models ran in real-time on a consumer-grade desktop computer, some required more processing grunt.

"Deep neural networks for guitar distortion modeling has been tested before, but this is the first time where blind-test listeners couldn’t tell the difference between a recording and a fake distorted guitar sound," said Välimäki. "This is akin to when the computer first learned to play chess."

Does this mean that professional gearheads will be swapping prized boutique rigs for plugins running in computer software in the near future? Maybe. Maybe not.

A paper on the Aalto project is available to read online, and the team has also published audio examples from the project.

Source: Aalto University

0 comments
There are no comments. Be the first!