24 ms·
Show HN: SmartGuitarAmp – Guitar plugin made with deep learning
- onion-soup 6y agoThe tone similarity is quite impressive.
- dlivingston 6y agoFor the guitarist interested in applying this plugin live, it would be worth looking into open-hardware Arduino-based guitar pedals [0]. [0]: https://shop.electrosmash.com/product/pedalshield-mega-kit/ https://shop.electrosmash.com/product/pedalshield-mega-kit/
- marketingPro 6y agoI'm ready for an esp8266 revolution. Arduino is outdated and expensive. Not that my wish will be reality, both academia and the last company I worked for used ardiunos for various tasks(non production). Not sure why, is it entirely the shape of Arduino is standard and can use Arduino based shields?
- detaro 6y agoThat revolution has pretty much happened already. There are giant ecosystems of non-Arduino systems out there. In many hardware DIY/maker circles you almost never see the traditional Arduinos anymore.
- bserge 6y agoFunny enough, esp8266 boards is all I can find in my country. Probably because it's cheap and versatile. I need to try it, but can't think of anything to use it for.
- qwertox 6y agoNowadays it's better to directly start using an ESP32, if you want to experiment. It has bidirectional I2S, (for example mic in, apply dsp, then out to speaker), is dual core, which would allow you to do the audio-related stuff on one core, and everything else (like communications) on another core. If you then see that you don't require all the good things an ESP32 has to offer, maybe because you only want to drive a LED strip via WiFi and read some sensor values, then you can step down and use an ESP8266
- marketingPro 6y agoWhy go for a esp32 if you can use i2c for reading many values and driving relays? Serious, because I am building a complex product and trying to keep costs down. Or are you referring to hobbyists?
- nitrogen 6y agoIf you're making something for audio you need I2S to talk to the DAC/ADC.
- qwertox 6y agoThe context of this thread is audio processing, so having I2S (Inter-IC Sound, with DMA) on board is really valuable, as is having an additional core. In my case, I can order a bundle of three esp8266 13.79€, or a bundle of three esp32 for 17.99€, both as the most lazy, fastest delivery option on Amazon. If you know that you don't need the extra features of the esp32, save yourself the 4.20€ (1.40€ per board) to keep the cost down. But if you're new to it and want to explore the possibilities of a modern µC, it would be a mistake to save in the wrong place. Even consider getting a LOLIN32 Lite (three for 19.99€), since the width of the board is smaller than the normal esp32 and esp8266 dev boards, which then lets you use both sides of a breadboard to plug in cables. All the other boards are so wide, that on a normal breadboard you will have to decide which side of the board you want to have a row of externally accessible pins, and on the other side all the cables must be plugged in below the board (search for "LOLIN32 Lite width breadboard") Saving on cost can also be a premature optimization. > Or are you referring to hobbyists? Any context that is not production. Development, Education, Hobby.
- sickcodebruh 6y agoThis is really cool, I’d love to try this out! Neural DSP, developers of what are arguably the best amp sim plugins available right now, claim that their upcoming floor modeler uses what they describe as “unique biomimetic AI technology” to create its sounds. I wonder how different their work is from this. https://neuraldsp.com/products/quad-cortex https://neuraldsp.com/products/quad-cortex I’m personally kind of over lugging an amp around and if live music ever comes back, I’m likely to sell mine and go with a modeler + power amp. The technology has just come so far and the convenience can’t be beat, especially if you do a lot of studio work or need to travel with your gear. A lot of demanding metal bands have switched away from tube amps and I’m already using plugins for recording as much as possible, so this seems like the next step.
- keyth72 6y agoThank you! Any feedback is appreciated if you get around to trying it out. If you look at the WaveNetVA repository from Aalto University that my plugin is based on, at least one person affiliated with that project works at Neural DSP. I suspect that the QuadCortex uses something like this (maybe a more advanced neural net) combined with other proprietary techniques and application specific hardware. They claim they can do the rig capture in 3 minutes, whereas the models I trained for the SmartAmp took a few hours on a GPU.
- bjelkeman-again 6y agoIt is really interesting to see this becoming more widespread. I am waiting for a delivery in the first batch of the Quad Cortex.
- sickcodebruh 6y agoIf real world reviews confirm the previews I’ve encountered, that’ll be the one for me. It really seems phenomenal.
- bjelkeman-again 6y ago
- _fs 6y agoI don't know, I'm not hearing the warmth of a tube amp. Sounds like other solid state emulator amps.
- tgv 6y agoIANAG, but it sounds like it's recorded through a speaker in the room, and the miking wasn't optimal. Now that might be the recording we watch, or the actual output of the plugin; no way to tell. But apart from that, it's quite good, IMO. (I Am Not A Guitarist)
- sosborn 6y agoAs a guitarist, it is important to understand that any time you tell a guitarist that the sound is digital, the guitarist will scoff and say it doesn't sound like real tube amp, regardless of what they said or thought before knowing it was digital.
- mhh__ 6y agoExactly the same with analogue synthesizers. A lot of the time they don't mean analogue when they say they don't want digital - they mean that they want knobs and switches to play with. I think some modern Moog's even have a digital stage in the signal path (I might be making that up but I have some vague recollection of it in a schematic) Making a synth sound good has very little to do whether it uses analogue or digital synthesis these days, but analogue equipment is (for the same price) less reliable, uses more power, less flexible, and more susceptible to being damaged by impacts (and keep in mind that you can't repair a lot of modern gear with SMD soldering experience anyway)
- klodolph 6y agoI will say that making digital synths that sound good is an incredibly difficult technical challenge. There are enough mediocre digital synths around that people can get turned off from them. Meanwhile, you can copy some old schematics out of a book somewhere and get something not far from a classic analogue sound. All of the mediocre analogue synths are dead and forgotten.
- aeleos 6y agoJust tried it out, I don't know if I would say it sounds like a $600 tube amp especially in high gain scenarios (it really didn't like my active pickups), for a free plugin it is definitely great. In comparison to the Archetype Nolly plugin from Neural DSP it doesn't really compare, but but for something thats free and can be trained at home on custom audio samples its pretty cool.
- keyth72 6y agoThanks for trying it, much appreciated. Love Neural DSP's plugins, they are the best as far as I'm concerned.
- biomechanica 6y agoMercuriall's plugins are amazing, too. Their Spark plugin and SS-11X (can't tell the difference on that one) are so unbelievably good. If this open source one can do what Spark can do.. I dunno.. I'd shave my beard.
- baptistelebail 6y agoYeah, totally. Their ReAxis plugin is a very solid plugin as well, delivering a full triaxis gear + some nice effects, the whole thing for less than $100. This thing is amazing and can deliver a high quality result for every kind of tones out there from clean to very high gain. What a time to be alive for a guitarist.
- Gollapalli 6y agoNeural DPS's plugins are great. I actually ended up ordering the real version of an amp I tried with a plugin. I never thought I'd say that! But their plugins are realistic enough that I can actually have some confidence that the real thing will be what I'm after.
- aeleos 6y agoNo problem, it’s great seeing guitar stuff pop up on HN, especially one that I can try
- romwell 6y agoWhat a clickbait title. Sounds (to whom?) like a $600 (which one?) amp using deep learning (why does it matter?). To expand on this: -$600 tube amp is not description of a sound. And a plug-in is not something you plug your guitar into to make enough noise for 300 people. "A plug-in sounds like a $600 speaker" makes almost as much sense as the title; by the sound of an amp people often mean the sound of the combo box which houses an amp and a speaker. -If this emulates a particular amp, the title should say which one -If it does a good job at emulating a particular amp, where's the side by side comparison? The video demo only shows the plug-in. -There are plethora of guitar amp plugins. What makes this one any different or better? -What's the significance of deep learning here? Plenty of ways to make this plug-in without deep learning. Why use deep learning? Etc, etc, etc. "An open-source guitar amp VST plug-in made with deep learning" would be a better title, even though the article leaves much to be desired.
- sbilstein 6y agoDon’t know why you are getting downvotes; these are exactly the questions any one with domain knowledge of playing electric guitars would ask.
- tgv 6y agoCome on... > (to whom?) Objectively. > is not description of a sound It is, as much as "piano" is a description. People are interested in emulating hardware in their audio software, and this emulates a guitar amp, and the price tag claims "it's a decent sound". > If this emulates a particular amp, the title should say which one No. > If it does a good job at emulating a particular amp, where's the side by side comparison? The video demo only shows the plug-in. Perhaps, but the article is about getting a free plugin. > There are plethora of guitar amp plugins. What makes this one any different or better? Nothing per se. It's just a new one. And it's free. It's an announcement. > What's the significance of deep learning here? That it's an uncommon way of modeling hardware, with its own merits.
- ysw0 6y agoIt's based on a Fender Blues Jr according to the vid. It sounds pretty good for a free plugin, but as you said there's tons of free VSTs and cheap solid state amps that can do better. I think it sounds good for a hobby project though!
- whiddershins 6y agoWhen I click through I see a git repo with a descriptive title. Why are we seeing this title in the HN submission?
- linux2647 6y agoProbably because the person who submitted the link is the same person who wrote the plugin
- keyth72 6y agoThis is correct.
- keyth72 6y agoI'm new to Hacker News and I might not understand how it's supposed to be used, can you explain what you mean so I can fix for future posts?
- detaro 6y agohttps://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html Among others: > Otherwise please use the original title, unless it is misleading or linkbait; don't editorialize.
- keyth72 6y agoAh, thank you for pointing that out. So if I wanted to use the title I have, I should write an article with that title. But even then it seems I wasn't in line with the quality of posts expected for this site. Will definitely keep that in mind for future posts.
- whiddershins 6y agoDon’t let any of this discourage you tho. This is a great contribution to the community. This site just has a preference to more technically descriptive titles.
- offtop5 6y agoWould be cool if you could run this thing on a raspberry pi and actually use it like an on stage amp
- TylerE 6y agoThat's unlikely to end well. You really need custom audio hardware - or at least dedicated commercial-grade stuff, for this sort of application. Latency is HUGELY important. 2-3ms can be felt, and most consumer level stuff uses buffers WAY bigger than that.
- bjelkeman-again 6y agoAs a reference the new kid on the block, NeuralDSP’s Quad Cortex uses 500 Mhz Quad-Core SHARC https://www.analog.com/en/products/processors-microcontrollers/processors-dsp/sharc-audio-processors-socs.html# https://www.analog.com/en/products/processors-microcontrolle...
- TylerE 6y agoYes, SHARC is pretty widespread. I'm not talking about the DSP though. You need an ADC on the frontend and a DAC on the backend.
- offtop5 6y agoThanks for the input. If I had 500$ or so could I make a digital bass pedal / tube amp ,?
- TylerE 6y agoMake, probably not.. But at that price point just get a used Helix LT which will be light years ahead of anything you can bodge together. This is one of those areas where both the hardware AND the software really matter. You can't touch the available hardware because of economies of scale - they're buying DSP chips, enclosures, power supplies, by the thousand at least. And they have multi-person full time teams working on the software.
- deeblering4 6y agoWhich tube amp(s) is it modeling? Its a bit like of like saying “free emulator behaves like $600 console”
- mhh__ 6y agoThe video says a Fender Blues Jr.
- easton 6y agoOT: looks like hn differentiates the domain for a GitHub site based on the user, guess it’s better to see what GitHub profiles have posted before.
- w0mbat 6y agoFor people who don’t play guitar, this is really commonplace technology now. Engineers worked out how to model the circuit behavior of any amp a while ago. Powerful valve guitar amps have really interesting effects on the waveform particularly when driven hard, but that valve technology is expensive, unreliable, bulky, and very loud when producing some of the most sought after tones. The most popular guitar amps now are all solid state, but the internal computer digitizes the guitar signal, then processes it in the digital domain, doing eq, applying the “sound model” of an old amp of your choice, also applying effects like chorus and reverb at the same time. Then the D to A convertor feeds the resulting signal to a neutral transistor power amp circuit to the speaker, at as low a volume level as you want. The BOSS Katana range or the Fender Mustang GT range are both very popular. They are good value, sound great, and can sound like anything from a Fender Twin Reverb to a Marshall stack set to 11. All without shattering your neighbor’s windows.
- MrBuddyCasino 6y agoIs latency an issue?
- joeel84 6y agoYes.
- mhh__ 6y agoLatency can be an issue if you are playing through a desktop computer or laptop, but its very manageable (for the most part) if you set up your buffers right. I have never ever noticed any latency issues with any physical hardware, even for equipment from years ago.
- o-__-o 6y agoA lot of hardware audio interfaces were usb2 which demands cpu time as the system has to poll the usb device rather than act on an interrupt. This caused very low latencies to basically never work without dropouts. I can distinguish a 20-30ms latency and above that it becomes too difficult to ply anything meaningful. Even my semi-pro tascam interface could only sample between 11-14ms. Faster cpus over the last 5 years and ultimately Usb3 solved the polling problem but is just now picking up traction in newer devices due to the usb-c connector.
- nom 6y agoAwesome project and pretty good results from a mathematical standpoint. The biggest challenge is to try and convince the tube amp enthusiasts that it sounds indistinguishable to the real hardware ;)
- dmacvicar 6y agoMost people, including guitar players, can't tell the difference between popular modelling devices and real tube amps when subject to a real blind-test. The of neural networks in this field should be measured against current modelling technology, not real hardware.
- trollian 6y agoThat's super cool. I've been both playing around with building analog pedals and thinking about how to cram various single board computers into a pedal enclosure for open source effect modelling goodness. How much compute does running the model use?
- lightlyused 6y agoI've been using guitarx, it includes some tube emulation, but I haven't really noticed much of a tonal difference between them. This looks interesting, but I run Linux so it is a no go.
- neiman 6y agoWindows only boohoohooo:(
- dharma1 6y agoHey, saw your comments a couple of weeks ago on the wavenetVA repo, this is looking great! Will try training spring reverb and analog tape, those have been pretty hard to model with DSP. Also, what do you think about diffwave? Only seen speech examples so far but seems an order of magnitude faster than wavenet. https://github.com/lmnt-com/diffwave https://github.com/lmnt-com/diffwave
- keyth72 6y agoNice, I’ll have to look into diffwave, have not heard of that. You may have difficulty training on reverb effects due to the nature of the WaveNet model. It works well for distortion, but for effects that extend past around 50 milliseconds like reverb or delay, it might not be able to capture it properly.
- fallingmeat 6y agoIn case any of you want to play with this on your mac and need a usb type-c interface to try this out! :) sorry for the plug but it was too relevant. Great job on the plugin! https://www.effintone.com https://www.effintone.com
- asimpletune 6y agoIt looks pretty neat, awesome form factor. It’d be cool to know more details without having to sign up though.
- fallingmeat 6y agoGonna fire off a kickstarter here in about a week and wanted to collect some followers as far ahead as possible! Promo is definitely not my jam! You're right, that page is totally worthless. Anyway, it's super simple. I just wanted a type-c version of what was already out there, and maybe for it to sound 'reasonable' at a sub $30 price point. It is only 16-bit 48kHz, but does a great job at dialing in tone, recording a quick riff, etc. Bottom line, it's about -90db noise floor and 80db SNR. Thanks
- asimpletune 6y agoThanks for the info, sounds like a great product.
- asimpletune 6y agoI was actually thinking that something like this in combination with a “reference” preamp would be the key. This is a similar approach that Townsend microphone’s have taken and they’re supposed to be absolutely fantastic. In any case, it’s total BS that waves, et. al, can charge hundreds of dollars per plug-in when at the end of the day it’s just software, and I was wondering to myself the other day why there isn’t more stuff in the open source space. I’m grateful to have seen this now, and I’d probably be interested in poking around this kind of stuff more. I have no idea how one actually goes about writing a VST or UA or AU plugin, but this is a great start, thanks!
- samirsd 6y agothis is very very good.
- manmal 6y agoInteresting that Kemper hasn’t been mentioned yet. They are a German company who make profiling amps, which basically consist of software able to profile an existing amplifier within ~2 minutes, and then perfectly replicating it. In an interview the inventor said that they have only on the order of 30 parameters which are configured during the profiling process, in which the device sends certain signals through the amplifier, and analyses the output. To make it stage-ready, it also has hundreds of effects and pedals (delay, wah, fuzz,...) on board. Those Kemper profiling amps are widely used in the music industry, from hobbyists to absolute professionals. There are studios who profiled their 100s of custom amps with different mic pickup settings, and now only use the profiling amp instead. Paul Gilbert has one, and loves it, and he is reaaally picky about sound. Apart from profiling your own amps, there is also a vast online library of profiles uploaded by community members. So if I want my amp to sound like Green Day, or just pick a certain amp like a Mark IV, I just search for that, load a matching profile into the profiling amp, and try it out. A plethora of parameters like gain can additionally be adjusted like on a normal amp - the profiling process somehow accounts for those parameters too.
- gianlucahmd 6y agoKemper is an amazing product, one of the few true innovations in guitar gear in ages. However, it still fails to model the response of a nice tube amp to dynamics. It gets close, but not quite there. This is why people like Paul Gilbert use it: he plays with so high gain and compression that dynamics would be basically flat.
- manmal 6y agoThat’s an interesting take for sure, and I think I read about this caveat elsewhere. Let me just add that Paul Gilbert has a very broad musical range - there surely is a lot of high gain shredding (at least in earlier days), but he‘s also doing amazing work with clean settings and acoustic guitars. His hearing is not the best though, so maybe subtle differences in dynamics escape him :) Personally I think that most problems people report about Kemper‘s sound stem from their speaker setup. If you use it for audio recordings or for loud stage setups, you most likely won’t be able to tell any difference. There are lots of blind tests out there and people often think that the Kemper is the real amp.
- qmmmur 6y agoI think that you can get lower latency and better sounding models with convolution. It's also quite fun to build your own impulse response library from things around your home, emulating the properties of your kitchen sink for example.
- oriolid 6y agoConvolution is linear, tube amps have many levels of nonlinear behavior.
- qmmmur 6y agoWhether or not it is nonlinear is another question. I'm saying that amp emulation using convolution is cheaper computationally, usually a lot more approachable and way more configurable. Ultimately people use this to make sound, so I'm commenting on the sound qualities that convolution affords which to me are good enough considering you've already gone digital to model something analog anyway.
- oriolid 6y agoThe thing is, convolution based guitar amp simulation is not going to sound like guitar amp, because convolution can't model the nonlinearity. You might get otherwise interesting sounds, but not what is usually considered the electric guitar sound.
- keyth72 6y agoWow, I am blown away by the feedback and discussion from everyone here. Just want to say that this is an extremely different experience from any other site I’ve been on, so thank you for that. Future goals for this project are to improve the audio quality and get community contributions for model training of different hardware. It’s very much still an experimental project, but I wanted to give an idea of what is possible using deep learning for audio applications.
- valdiorn 6y agoThis will get lost in the pile, but I've done some tests with a dumb-as-a-brick neuralnet, 5 layers of 128 nodes, reLU, and the thing shows INCREDIBLE quality. I build VST plugins as a hobby, and have done several amp and guitar effects plugins. I also build and design tube amps. I was very pleasantly surprised at the quality I was getting, and it was simple enough to run in real-time using pytorch (well I exported the model to onnx and used a c++ lib to run it). I was on my way to making a little startup company around the idea when covid hit and priorities radically changed
- copperx 6y agoWould the neural net be useful for modeling other hardware? Or is it configured specifically for guitar amps? If so, would it be able to model something like a hardware compressor?