Yamaha DX7 chip reverse-engineering, part 4: how algorithms are implemented | Hacker News Reader

Yamaha DX7 chip reverse-engineering, part 4: how algorithms are implemented

righto.com

202 pointsTomte48 comments

This article series is mind blowing to me. The engineering is fascinating. It's like they made another logical analog abstraction layer over the digital electronics. Yamaha doesn't get mentioned much here, but between how they can bring the experience of their instruments, motorcycles, and sound equipment to people everywhere, there is an understated beauty in what that company does. They only seem to make things that are wickedly fun and life's great pleasures. Also, this: https://en.wikipedia.org/wiki/Yamaha_CX5M

Yamaha is an incredible company. I love seeing the tuning forks on a Yamaha motorbike and thinking about all of the work that has been done there.

Here is a very good WebAudio implementation of DX7, so you can hear what it actually sounds like: https://www.webaudiomodules.org/wamsynths/dx7

Doesnt work on iphone

It's not mine, so I can't help. But I know iPhone / Safari has pretty terrible support for lots of things I've done having to do with midi and WebAudio, and they don't allow Chromium to run. Not sure why people consider that ok, but such is life.

Works on mine..

You might have to toggle your iPhone’s silent mode switch, as it mutes webaudio.

My DX7 is still one of my favorite synths, sonically speaking. I had no idea how intuitive yet novel the circuitry was until I began reading this article.

In particular, I found it suprising that the sine waves were generated by indexing a wave table - you'd think a sine wave would be easy enough to generate without a table, but the way it implements phase distortion ('FM') on those sine waves by simply offsetting the index into that table is quite interesting. Sorta like having different mathematical functions for 'i' in a for loop.

In DSP terms direct calculation of sine waves is incredibly expensive. Even more so with 1980s 8- or 16-bit fixed point arithmetic without a coprocessor.

Table lookup was inherited from the first computer music software. It was the only way to create any kind of waveform at reasonable speed.

When you have a table, modulating the lookup index is almost an obvious trick. But the Chowning/Yamaha implementation worked so well musically that it took it from a curiosity to a best-selling product.

A different take on it might not have done that.

FM owes a lot to sound developer Dave Bristow. He made sure the DX7 had usable sounds. According to him, the original prototypes had tens of harpsichord-like patches and not much else.

Intriguing to imagine the first time the designers heard the sound their creation could make (presumably there was a TTL or bit-slice prototype before the VLSI implementation). They're sitting around the lab, someone plinking away on the keyboard. One of them says "you know, you could use that sound on a pop record and I bet it would be popular for at least a decade".

The story behind FM synthesis is pretty interesting. A Stanford music professor, John Chowning, came up with the idea in the 1960s and patented it. Stanford didn't think this was what a music professor should be doing and fired him. Meanwhile, Yamaha licensed the patent from Stanford, paying millions of dollars and making it Stanford's most lucrative patent at the time. Stanford changed their mind about Chowning and hired him back, making him a full professor and then department chair.

For more information, see: https://priceonomics.com/the-father-of-the-digital-synthesiz...

The early FM pieces weren't very pop. https://www.youtube.com/watch?v=988jPjs1gao

You may think it sounds crude, but pieces like these took hours of expensive mainframe (PDP-10) time. There wasn't a lot of opportunity for careful sculpting of fine details.

You could easily synthesize something like this in real time now. But not many people do, which I think is a shame.

If you can read French, there's more background here: http://brahms.ircam.fr/analyses/Stria/

That clip sounds like something straight out of the soundtrack from Close Encounters of the Third Kind.

The synth used on the movie was an ARP 2500. Fascinating machine.

https://reverb.com/news/a-brief-history-of-the-arp-2500

hours of expensive mainframe (PDP-10) time

Equivalent today to milliseconds of processing in an Apple Lightning cable.

> If you can read French, there's more background here: http://brahms.ircam.fr/analyses/Stria/

If you can't read French, use Google Translate. It handles this article extremely well.

You could easily synthesize something like this in real time now. But not many people do, which I think is a shame.

I've kind of been following very loosely along with this DX7 series with some (not DX7) FM and phase modulation in my own code. If you're running Linux, you can try something like this in the WIP graph-synth branch[0] of my mb-sound Ruby app/gem (feedback welcome on the DSL syntax):

    play 300.hz.pm(
      0.1.hz.triangle.at(1) * tone(300.constant * 1.618) +
      0.04.hz.triangle.at(1) * tone(150.constant * 1.618)
    ).forever

[0]: https://github.com/mike-bourgeous/mb-sound/tree/graph-synth

doesn't sound the same though, those early Yamaha FM synths has a sound all on to themselves.

those early Yamaha FM synths has a sound all on to themselves

For sure, though the parent comment's video link was presumably rendered by software on a PDP-10. I've spent the last week of my spare time trying to write code to make a bass sound that reasonably approximates Lately Bass, and I'm close, but since I'm not even trying to emulate any of the quirks of the Yamaha FM synths it doesn't ever sound quite right.

To be clear - I didn't mean it was a shame that people weren't specifically emulating FM, but that people weren't experimenting with weird synthesis techniques to make weird non-pop music with sounds no one has heard before.

It's the last part that's still really difficult, possibly even harder than it used to be. And that's even with the insane processor power we all have now.

I guess movie soundtracks are the closest to that in widely distributed music, or artists like Aphex Twin.

There's an infinitely vast space of potential sounds, and an infinitely long fractal knife edge of interesting sounds, and it kind of makes sense that finding unexplored areas of that fractal edge between boring and incomprehensible gets harder over time.

speak for yourself! hehe

https://soundcloud.com/cassilda_and_carcosa/ontologies-tape-...

c15 on the tx81z, you'd have to model the 12 bit DACs and the steppy-nees of the wave shapes. also the CPUs in the tx81z were slow and some calcs would lag a bit which contributed to its sound... try sending midi data to the TX81z while sequencing it, your pattern with start swinging!

TX81z is one of my favorite synths.

> Stanford changed their mind about Chowning and hired him back

Somehow I feel that if this happened today, no company or university would have the spine to do this, and it would rather turn out as an expensive legal battle if he insisted on getting his part of the patent pake.

Great read. Images aren't loading on that link but wayback machine to the rescue: https://web.archive.org/web/20150323222119/https://priceonom...

My favourite telling of the DX7's amazing history is in this video by MadFame: https://www.youtube.com/watch?v=sXt_NXjc7oY

His Youtube channel has some absolutely amazing patch tutorials also.

Author here if anyone has questions...

Just want to say that I am incredibly thankful for this series, its amazing to see how one of my favorite synths works not just in theory (FM/PM) but also on the chip level. Thanks!

I feel like there was a missed opportunity with the title, why the departure from using Roman numerals and call this article part IV?

A missed opportunity in what way? Is "IV" better than "4"?

The previous article in the series used III to denote part 3. IV would be next. I guess part 5 could be titled part Five, or maybe Go since yamaha is a japanese company.

Thank you for a good series of articles and some great insight into the inner working of FM synthesis in general.

I have never owned one of the original Yamaha's, but i have been using the awesome Dexed VSTi plug in, that is modeled after the DX7 for several years now.

And I am planning on purchasing a Korg Opsix for its FM capabilities.

But for someone that has only been on the learning-by-experimenting end of FM synthesis, your articles are a great insight into the theory behind it all.

How much time does reverse engineering this kind of chip take?

It takes a while. I got the chip on Nov 1 and have been working on it since then. (Although it's not the only thing I've been doing.) The process is a combination of taking die photos, tracing out circuitry, understanding the circuits, doing background research, and figuring out how to explain the chip in blog posts.

I’ve never been a huge fan of FM synthesis as I’m mostly reminded of bad electric piano sounds that came out back in the 80s. Having owned a tx-7 I knew it was capable of more than that but then there was the whole issue of programming it that kinda sucks.

That said seeing the tech and how things were implemented has me loading up the Arturia DX7-V and tinkering around with it.

Have any other synths or effects on your list to look at?

Once you got past the bad piano sounds, it was in way too many arcade and home video systems to keep track of:

https://en.wikipedia.org/wiki/Yamaha_YM2151

One of my favorite pieces is Brian Schmidt's music for Swords of Fury, a Williams pinball machine from 1988.

https://www.youtube.com/watch?v=RMIp5nG-C3o

On the subject of favorite FM pieces, this 1992 track from Jill of the Jungle (rendered by one of Yamaha's OPL2 chips on a Sound Blaster) burned its way into my brain as a kid. Still a fan: https://www.youtube.com/watch?v=8pcxdYvp8uQ

FM can be used for some incredible atmospheric pads too, like the one at this timestamp: https://youtu.be/8SDlp5jlUqc?t=416

Lovely work - i'm wondering if you now have enough info to build a VHDL model from this? It would be very cool to run an emulation on an FPGA.

My interest in these instruments goes back a long way - I reverse engineered my DX7 to build a VST instrument around 2001 or so. My reverse engineering was 'black box' with putting test patches and midi through the instrument to build a model, but was good enough to produce patches which I couldn't tell from the original except at very extreme settings where the different in sample rate became apparent (e.g with aliasing frequencies being very obviously different between the two).

Actually, thinking about it, I also did the same with the TX instrument to produce a simple model written in SOUL a few years back. There's a LatelyBass patch here based on this - https://soul.dev/lab/?id=LatelyBass

It has been done for OPL3 https://github.com/gtaylormb/opl3_fpga

I have a suspicion emulating in software specific Yamaha hardware implementation would not only lead to bit perfect reproduction, but also faster emulator. I wonder how slow of an MCU would one need, 20MHz AVR? 30MHz STM32?

Ooh, that's an interesting project, definitely lots there for me to learn about. Thanks!

I'm not sure if these have been posted before, but the progress notes from Yamaha are interesting - https://ccrma.stanford.edu/~aj/archives/docs/all/051.pdf

https://mmontag.github.io/dx7-synth-js/

Looked for an in-browser DX7 synth, found this!

I'll just loudly repeat my call that it would be great if someone put this much effort into taking apart the Roland D-50. Just sayin'...

...please?

I worked in a music store back then. The DX7 was KING for sure but we were a Roland Shop and the D-50 was pretty F'n awesome. I don't think it's apples-to-apples though. Wasn't the D-50 less of a synth and more of sampling board?

Yes and no. I guess the best way to describe it is that it was more of a half-step towards wavetable synthesis in the sense that it utilized ultra-short samples of the attack and sustain phases of the envelope to achieve something with greater fidelity than nominally possible with FM synthesis alone, but well short of what you could achieve with, say, a Korg Wavestation. It was an interesting product of its time.

I did a huge amount of D50 programming. The actual digital synthesis was not particularly impressive (limited architecture and weak filters). Multiple voices ('partials') could be layered in interesting ways - quantity over quality! The synth sounded great because because of the samples and the built-in effects.

You can listen to each sample if you try programming a patch. Some are short 'transient' samples, others are looped. The 'chipmunk' effect is really evident.

The D50 was the first of the new wave of 'S&S' ('sample and synthesis') synths that had nothing of the depth and mathematical interest of FM synthesis.

I have a DX7 I might part with...

Reply on news.ycombinator.com