Flashback #7.1: The H949 Harmonizer®

“Ding Dong! The Glitch is Dead”

The success of the H910 was gratifying, especially considering its limitations; for one thing, you could only do so much with 100 milliseconds of delay! We were already thinking about a new, improved model even as the first H910s shipped. At the time—the mid ‘70s—IC technology was sprinting ahead. The 4k-bit RAM chips in the H910 became ‘old news’ once 16k-bit chips were available. And, by 1977, logic ICs had progressed to doing simple but fast arithmetic, which enabled a host of unheard-of effects.

The H949 benefited from the rapid pace of IC development with improved audio specs and much more. It represented a major advance in the very notion of an effects box and introduced the terms algorithm, random, and micropitch to the audio lexicon. Our marketing message was simple—“more of everything.”

Our ad cheekily called out some of the novel features:

The Glitch’s Tale

The long list of “more of everything” included longer delays, radical new features, and better audio specs. But, for this Flashback, we’ll focus on the ‘devilish pitch change glitch’ that plagued the H910 and the method that we came up with for smiting it.

Why was taming the glitch so important? Didn’t some people love that devilish glitch? Yes, perhaps some did, but most simply tolerated the glitch while appreciating that the H910 opened up a new world of sonic possibilities. Others had the hope that it could be used to help solve a sticky problem: ‘pitchy’ vocals. On that score, it fell short for two reasons. First, it was difficult to dial in small, precise pitch ratios, and second, the random glitch made for hit-or-miss results.

The H949 was the first pitch-change box designed to be a tool for tuning. It had the necessary fine resolution, as well as the ability to analyze audio in real-time and make decisions that avoided audible glitches. Engineers welcomed this new capability and found that while monitoring a problematic track, they could twist the big knob—at the right time and by just the right amount—to bring a wandering pitch in tune. Bear in mind that, in 1979, autotune was still more than a decade away. This process was a hands-on, real-time performance. Engineers discovered that the H949 could bail them out. Here’s a case in point:

The path to solving the glitch problem took a number of twists and turns. If you’re game, come along with us as we travel Nerd Boulevard! Alternatively, flash forward to Flashback 7.2 for a deep dive into the H949’s many groundbreaking features.

Why a Glitch?

Let’s explain what caused those devilish glitches in the first place. The H910 was built with the earliest ICs—simple logic gates that could do little more than calculate a couple of memory addresses for each audio sample. It did its best to ‘smooth over’ the discontinuity that occurs when a pitch changer’s delay is forced to suddenly ‘jump’ by several milliseconds when the delay either gets to zero (increasing pitch) or becomes too long (decreasing pitch). The H910 employed essentially the same method as human tape editors who used razor blades and a cutting block to splice tape at a 45-degree angle. The audio segments would, in essence, crossfade. Rather than risking a hard splice (and possible loud ‘click’), the crossfade smooths over any discontinuity. And yet, while the H910 employed the same crossfade splicing method as editing tape, there are two significant differences:

Over the course of a 3-minute song, there might be as many as a dozen tape edits/splices. A real-time pitch changer, depending on pitch ratio, may make hundreds of splices over the same period.
The tape editor can listen to the audio and ‘decide’ where to splice. The H910 could neither listen nor decide. Its splicing method was aesthetically agnostic with unsurprisingly uneven results.

We called the artifact that resulted from the pitch changer’s random crossfading the glitch. The experience with tape editing made it clear that the way to handle the glitching problem was to do what a human tape editor does: Find ‘good’ places to splice. If only the electronics could compare two portions of the signal so that, when a splice was necessary, it would occur between two points on the track that were similar to each other. Finding such a similarity uses a process called autocorrelation. So, the ‘solution’ was rather clear. In fact, across the pond, a couple of ex-aerospace engineers, Mark Crabtree and Stuart Nevison at AMX, reportedly had arrived at the same ‘bleeding obvious’ conclusion.

While, theoretically, autocorrelation was the way to go, the practical reality was that until the late ‘70s, ICs weren’t up to the task of ‘processing’ audio in real-time. Autocorrelation of a real-time audio signal requires many thousands of multiplication operations per second. ICs capable of doing this arithmetic quickly were still years away. So, we decided to hedge our bets by providing our new model Harmonizer® special effects unit with the option to select one of two methods for pitch change. Anticipating the day that ICs would be up to the task of real-time analysis and decision making, we designed the H949 with the ability to add new, ‘intelligent’ hardware that could make smart splicing possible.

The Algorithms

The H949 gave the user two methods to mitigate glitching, which we dubbed ‘algorithms.’ The User Manual introduced the word ‘algorithm’ to the audio community, defining it as “a precise, describable process which acts upon or modifies inputs in a specific manner.”

Algorithm #1 was new, yet simple. It created slow, gentle crossfading that eliminated hard glitches, but caused a ‘swimming’ effect at extreme pitch ratios. When the first H949s shipped, Algorithm #2 was similar to the H910’s method and had similar, random results. The plan was to offer a hardware upgrade for Algorithm #2 when newer, faster ICs would make intelligent splicing, and hence de-glitching, practical.

Intelligent electronics?

Eliminating the glitch required more than just the intelligence for signal analysis and decision making; it also had to perform the complicated signal memory addressing required for de-glitching, and to make those computations at audio rates.

Real-Time Audio: It’s a Matter of Time

Early ICs were not only vastly simpler than today’s devices, but also much, much slower. Real-time digital audio has a fundamental constraint in that processing has to be completed on a sample-to-sample basis. Assuming a 50 kHz sample rate, the analog to digital converter spits out a new sample every 20 microseconds and the processing must keep up with this never-ending flood of samples, forever.

Today, processing audio in real-time is no longer an issue, and in fact, it hasn’t been for about 3 decades. The microprocessor in your cell phone can run audio algorithms far more complex than that in the H949 while it’s playing video games. But in the early IC era, dedicated logic circuits were needed to get the simplest tasks done, e.g., to perform a couple of additions to calculate an address to keep up with real-time.

Bit Slice ALU for Addressing Memory

Enter Advanced Micro Devices, nowadays a credible competitor to Intel. The first Arithmetic Logic Unit (ALU) capable of computing memory addresses fast enough to keep up with real-time audio was their ‘bit slice’ AM2901. What’s ‘bit slice’? Each chip could add, subtract, and store only 4-bit data (a “nibble”) but multiple chips could be combined to perform simple arithmetic on a digital number of any length. To address the 16K memory of the H949, addresses had to be at least 14 bits long, so our ALU used four bit-slice chips to create a 16-bit number. Here’s a photo of the four chips that were the heart of the ALU:

Microcode

The high-speed ALU of the H949 made it possible to compute—hold on to your hats—16 instructions to handle splicing and 16 memory addresses in every 20-microsecond sample period! In other words, the ALU’s entire ‘software’ program consisted of a total of 16 lines of ‘code.’ Here’s a sample of Tony Agnello’s handwritten instructions for handling splices:

With this fast addressing capability, the H949 could offer new features like random delay, reversed pitch change, micropitch, and, if only some intelligence could be added, smart splicing. Adding the intelligence necessary to de-glitch seemed years away…

…But then we got lucky!

The Mystery Chip

Here’s the H949 with its top board flipped up:

Take a close look at the add-on circuit board, the LU-618; dubbed the Lupine Board:

The Lupine board harbored a secret. Note that the part numbers were erased on the two big chips:

Why the mystery? Well, as mentioned, we had gotten lucky and we didn’t want our competitors to ‘share’ our luck. Thanks to the military’s need for ‘impossibly’ fast processing, some clever souls at a company called Reticon came up with a way to use special analog CCDs (Charge Coupled Devices) to analyze signals in real-time nearly a decade before the first DSP chips were available! This obscure and one-off technical advance made it possible to create what was arguably audio’s first intelligent, real-time audio processor. The LU-618* ‘de-glitch’ board with the Reticon chip was offered as an option ($740 in 1979 ≈ $3,000 today!) and well worth the bucks when one considers the cost of studio time—or if the band had left the studio, the impossibility—to re-take pitchy vocals.

The H949: At the Dawn of Intelligent Electronics

The H949 was at the forefront of smart electronics. It was able to analyze audio in real-time and make decisions based on that analysis. It was used to create new sounds and to correct pitchy tracks. We’ll close with the original draft of our initial H949 magazine ad. Can you find the typo?

Here’s the data sheet for the Reticon chip, the charge-coupled (switched analog) correlator IC designed for military applications (radar). This was the missing piece enabling the design of the first de-glitched Harmonizer. The H949 used autocorrelation to analyze the audio in real-time and, based on that analysis, intelligently select the splice points.

In 1979 the H949 appeared at the Audio Engineering Society (AES) Convention in NYC, here’s the original press release:

*Why LU-618? Heather Wood, late of Dolby, was our marketing department at the time and a Monty Python fan, as were we all. When the subject of selling this upgrade was being discussed, she dubbed it the “Lupine” board and thus it became. The origin of the numerical suffix is lost to history.