I'm working on a project to equip model locomotives with sound boards. I'm in the process of designing the board at the moment, and the idea is to allow users to load their own sound files onto an SD card plugged into the board.

Conventionally, model locomotive sounds are collected from high-fidelity microphones placed on and around the real engine in question. The engine is started up then put through idle and all of the different notches, as well as dynamic braking, horn and bell sounds, etc. This practice is very expensive because you have to find a willing (usually small) railroad or museum, pay for travel expenses, and diesel fuel ain't exactly cheap at the volumes these engines go through. Secondly, newer engines are hard to record because railroads aren't exactly in the business of letting hobbyists tape microphones all over their money making machines. As such, the main cost for a sound board comes not from the circuit's BOM cost, but from the effort required to get sounds from locomotives.

What there's plenty of are YouTube videos of amateur rail enthusiasts taking videos of locomotive sightings at close(ish) proximity, including startups and shutdowns. My question is - is there a way to take a bunch of different audio recordings of the same engine, remove noise and the doppler effect, and from that create a profile that can be used to simulate what the engine might sound like at different throttle notches? Is machine learning the right tool for this?

  • 1
    $\begingroup$ Deep learning has been used to create a super-resolution photo from many low-resolution photos (arxiv.org/pdf/1808.03344.pdf), and also adding things to the images (e.g. make a frowning person smile). I don't know of an application that creates a super-resolution audio from many low and noisy resolution audios, and then predict the audio at different notches/braking/horn/etc, but I don't see why it can't -- it just hasn't been done before (that I know of). Do you plan on developing a model of this yourself? $\endgroup$ Dec 11, 2020 at 6:28


You must log in to answer this question.

Browse other questions tagged .