Home›Articles›Hartmann Neuron: The neural-network synthesizer that wo…
Articles

Hartmann Neuron: The neural-network synthesizer that won its awards before it shipped

Hartmann Neuron: The Hartmann Neuron promised something no keyboard had offered: you feed it a sound, a neural network studies it, and you play the resulting…

Bernardo de Azevedo
Bernardo de Azevedo
· Updated
Hartmann Neuron - illustration

The Hartmann Neuron promised something no keyboard had offered: you feed it a sound, a neural network studies it, and you play the resulting model as if it were a living instrument. It collected innovation prizes, and by late 2005 its maker’s website was gone.

That arc has a familiar shape in synthesis. A brilliant idea reaches the market before the product around it is ready. But the Neuron’s case is more specific than a generic tale of overreach, and the details, from a 29-minute file transfer to a bankruptcy notice posted on a synth blog, show exactly where ambition outran engineering.

Table of Contents

  • Hartmann Neuron: a sampler that did not want to be a sampler
  • The people behind the machine
  • Resynthesis, neural networks and what the marketing promised
  • What Sound on Sound found in 2003
  • The hardware that behaved like a computer
  • A cheaper Neuron, and a website that went dark
  • Who still plays it
  • Final perspective

Hartmann Neuron: a sampler that did not want to be a sampler

The core idea of the Hartmann Neuron was that a sample is a dead end. A sampler plays back a recording, with a bit of filtering and looping. The Neuron instead analyzed a recording, built a model of it, and resynthesized sound from the model. Wikipedia’s account says the model could then be adjusted through musical properties, such as the body size and material of the imagined instrument.

The instrument was designed and built by the industrial designer Axel Hartmann under the name Hartmann Music in Germany, over a development period that Wikipedia gives as 2001 to 2005. Its infobox lists 2003 as the date. Launch price was about $5,000, and Synthtopia gives a manufacturer’s suggested price of $4,990 or €4,890. In Britain, Sound on Sound put it at £3,499.99 including VAT.

The people behind the machine

Axel Hartmann is better known as a designer than as an engineer. In his NAMM oral history, recorded in January 2019, he is described as a music industry designer with more than three decades of experience, who began working with Waldorf Music soon after finishing his design studies. His Waldorf Wave won the Red Dot Design Award in 1992. The same page credits him with releasing the Neuron around the turn of the millennium and notes that film composer Hans Zimmer valued its cinematic sound.

The synthesis technology came from outside. Wikipedia credits Stephan Bernsee of the software company Prosoniq. Bernsee, born Stephan Sprenger in 1971, founded Prosoniq and is described as the principal developer of the Neuron’s synthesis technology. Sound on Sound’s review calls the method Multi Component Particle Transform Synthesis and names Stephan Sprenger as its designer. Taken together, those sources point to a division of labor in which Hartmann shaped the instrument physically and Prosoniq supplied the engine inside it.

Resynthesis, neural networks and what the marketing promised

Bernsee’s own blog gives the clearest account of the idea. Classical resynthesis, introduced in the 1980s, rebuilds a sound as a sum of sine tones at different frequencies and amplitudes. It becomes unwieldy with hundreds of tones, and sine tones handle short, impulsive sounds poorly, often producing a diffuse quality called phasiness. The Neuron instead adapts its basis to the sound being modeled, which, he wrote, allows faithful reproduction and wide modification.

The blog describes the model creation as a pattern recognition process based on artificial neural networks. Because the analysis is computationally heavy, it was done ahead of time in a separate program called Modelmaker, not while playing. A complexity parameter limits how detailed the models become, and complex material such as drum loops is represented by a time-variable morph between models.

Hartmann’s marketing leaned hard on the neural network. Sound on Sound’s reviewer found that, as tested, the Neuron did not select its own parameter sets in the way the marketing implied. The user chose the settings.

What Sound on Sound found in 2003

Gordon Reid reviewed the Neuron for Sound on Sound in August 2003, and the verdict was a careful one. He called it something of a mixed bag and an expensive work in progress, and said he could not wholeheartedly recommend it. He thought the factory sounds were flat without effects, and that many models had loop artifacts and octave jumps.

He also praised what worked. The real-time controls, especially recorded joystick movements that Hartmann called stick animation, gave a degree of control he said was unavailable elsewhere. Effects made pads and textures more interesting, and the machine did not crash during the review. He thought it suited sampling, mangling, sound effects and foley.

The most pointed remark concerned timing. The review questioned the awards the Neuron had already received before the first units shipped. Synthtopia, writing in 2009, makes the same point in its own words: the awards were presented nine months ahead of delivery, and the instrument was more ambitious than the company’s resources allowed.

The hardware that behaved like a computer

A good deal of the Neuron’s trouble came from being a computer in a keyboard case. Wikipedia describes a PC base running an early version of Gentoo Linux with a custom synthesis application and a custom sound card, in a heavy aluminum chassis. The Sound on Sound review lists a 20 GB internal drive, a boot time of about 40 seconds that the reviewer found unsuitable for live use, and a fan he considered annoying in a studio.

Synthtopia’s specifications give the drive as 10 GB, so the figure differs between sources.

Transferring the factory library of about 300 models, roughly one gigabyte, took 29 minutes. Models were also much larger than their sources: a 268 KB AIFF file became 1.1 MB at the lowest complexity setting and 3.0 MB at the highest. Polyphony was modest at the time, with voice stealing starting around the sixth or seventh note, and the audio inputs were not working yet.

Creating your own sounds involved a separate computer, a USB-to-Ethernet adapter, network configuration and an FTP client. The reviewer also noted that models did not crossfade across the keyboard, which produced audible jumps. Wikipedia adds that the limited display made adjusting tonal qualities a slow trial and error process.

A cheaper Neuron, and a website that went dark

In 2005, Hartmann released Neuron VS, a software version for Mac and Windows. Sound on Sound reviewed it in April 2005 at £499.99 including VAT, against £3,500 for the keyboard. The reviewer said it delivered most of the synth engine for about a seventh of the price, but it was very demanding on the host computer, needed a USB dongle called Nuke and lacked surround output and multitimbrality.

The company’s end came quickly. On December 5, 2005, the blog MatrixSynth reported that Hartmann Music’s servers had gone offline. A reproduced notice from Denis Goekdag, a former Hartmann support employee, cited the company’s ongoing bankruptcy filing and said no further tech support or downloads would be available. He offered free support through his own company, Surround SFX, which he said was not affiliated with Hartmann.

Wikipedia attributes the insolvency to a dispute with the distributor, though the MatrixSynth post does not name a cause. A commenter added that new Neuron and Neuron VS software was available through Bernsee.

Who still plays it

Wikipedia lists musicians who have used the Neuron, including Peter Gabriel, Manfred Mann, Hans Zimmer and BT, and says it remains in use mainly in film music. Synthtopia’s comment thread gives the other side of the story. One reader borrowed a unit for a week and found it sounded fine but was not fun to play, saying it booted slowly and resembled a PC. Another owner of the software version praised its didgeridoo sounds.

Such split verdicts match the reviews from 2003. The Neuron’s sounds were distinctive, and its workflow made them hard to reach.

Final perspective

The Hartmann Neuron deserves to be remembered less as a failure than as an unfinished argument. It tried to replace the sample with a model, to let a musician reshape an instrument’s body and material instead of a waveform, and to bring pattern recognition into a keyboard. The idea was sound enough that its engine survived the company as software.

What sank the hardware was the distance between the claim and the experience. Awards arrived before shipping, the user interface was slow, and the price was that of a studio centerpiece for an instrument its first reviewer called a work in progress. For anyone building tools around machine learning today, the Hartmann Neuron offers a plain lesson: the demonstration can be convincing while the day-to-day instrument is still missing.

That is why the Hartmann Neuron story keeps drawing musicians and historians back to it.

Few chapters of synthesizer history show as clearly as Hartmann Neuron how far a single idea can travel.

Looking back, the Hartmann Neuron episode suggests that the most unusual stories are often the most instructive ones.

For anyone tracing the history of electronic instruments, Hartmann Neuron is a name worth remembering.

Related stories

Sources

Bernardo de Azevedo
Written by
Bernardo de AzevedoFounder & Editor

His connection with music began at age 6, in the 1980s, when his father introduced him to Jean-Michel Jarre's Rendez-Vous on vinyl. He works professionally in the legal field, while synthesizers became his space for abstraction and creative exploration. Founder of The Synth Source.