Scat voice expansion pack

I’m pleased to release version 1 of my jazz scat voice expansion pack for Yamaha PSR-S950 and PSR-S750 arranger workstations. The expansion pack has five PSR voices which let you create “Take 6” style, a cappella arrangements and other kinds of jazz voice performances. Give the MP3 demo a try!

Four of the PSR voices are individual syllables: DOO, DOT, BOP and DOW. The DOO syllable is looped and let’s you create sustained chords for backing. The DOT, BOP and DOW syllables are short and provide scat-like expression. All four syllables are combined into a velocity-switched voice where you select and play one of the syllables based on how hard you strike the keys (i.e., MIDI note velocity). You will need to adjust touch response (and practice!) to get the most playable and musical result.

Here is a link to the expansion pack file. You need to download and UNZIP this file, then install the YEP file by following the directions in the Yamaha PSR-S950/PSR-S750 Owner’s Manual. See the section titled “Expanding Voices”.

I am also releasing the multi-samples that I used to create the expansion pack in case you would like to create a scat voice for your own synthesizer or software instrument. If you are curious about how I created the expansion pack voices and the samples, please see this blog post.

Both the scat voice expansion pack and the scat voice samples are released under a Creative Commons Attribution 4.0 International License.

Creative Commons License
ScatVoices and ScatVoice samples by Paul J. Drongowski are licensed under a Creative Commons Attribution 4.0 International License.

You are free to use the expansion pack voice or samples (even for commercial purposes) as long as you provide a link to from your own web site AND/OR explicitly credit me in your creative work, e.g., “Scat samples/voice by Paul J. Drongowski”.

Sampling “scat”

In this post, I describe the process and tools that I used to capture samples for my jazz scat voice. I will eventually release the voice (for the Yamaha PSR-S950 workstation) and its samples under the Creative Commons attribution license. I’m not the best singer, so I’ve had to rely on technology as much as possible while still producing a musical result. I want to emphasize that I sang, edited and produced all of the samples and the voice patch; it is original work.

The jazz scat voice is inspired by the (in)famous “jazz voice” patch found in Roland keyboards. The Roland patch is based on samples from the Spectrasonics Vocal Planet library by Eric Persing and Robby Duke. Their work was clearly influenced by Take 6 and other contemporary a cappella artists.

My patch uses four multi-samples where each multi-sample is a particular syllable taken over 12 (or so) pitches. The multi-samples cover the natural range of the human voice from F3 to F6 where C5 is middle C. The four syllables are: DOO, DOT, BOP and DOW. The DOOs are long, looped samples that provide a musical bed or harmony. The remaining three samples are short one shots suitable for melody, punctuation and accents. The DOW syllable falls.

The basic patch design is summarized in the following table.

Syllable Type Vel low Vel high Gain
DOO Loop 1 89 0 dB
DOT One shot 90 105 -3 dB
BOP One shot 106 119 -6 dB
DOW One shot 120 127 -9 dB

The table shows the MIDI velocity range to each syllable (multi-sample). It also shows the relative gain for each syllable. The gain decreases as velocity increases in order to maintain a more consistent volume level as the keys are struck harder to trigger the one shots.

At a strategic level, the sampling production process consists of two major steps:

  1. Capture a natural voice sample for each syllable and pitch. These natural voice samples are the formants to be used in the next step.
  2. Capture a vocoded sample for each syllable and pitch while playing the appropriate formant sample through the PSR-S950 vocoder.

This process produces scat syllable sounds that are consistent, pitch accurate and in the case of the DOO syllable, loopable.

Here’s a run-down of the practical problems that motivated this approach. My voice is an untrained baritone. It cannot possibly cover the F3 to F6 range without hysterical noise and possible voice damage. As I discovered, it is nearly impossible to sing pitch accurate short syllables such as these without proper training! I needed to find a method that would give me a consistent and pitch accurate sound across the desired range of pitches. This is a greater challenge than I originally anticipated and a lot of experimentation led to the two-step method. It took about 3 weeks to find the method and then a further two weeks of production work.

Now, the details.

I used a Roland Micro-BR digital recorder to capture both natural voice and vocoded samples. This little wonder is great — easy to use, fast and above all, quiet. For natural voice, I sang into a Shure PG-81 condenser microphone feeding an ART TUBE MP preamp. The TUBE MP is a really Swiss army knife providing phantom power for the PG81, a little bit of tube warmth, and conversion from XLR to a line level audio signal. The output of the TUBE MP is connected to the Micro-BR. For vocoded voice, I connected the line level mono output of the PSR-S950 to the Micro-BR. In both cases, all Micro-BR input effects are disabled and gain staging is established before hitting the RECORD button.

Formants are captured and produced in the following way. I sang each syllable multiple times at each of the desired pitches while recording to the Micro-BR. The pitches cover the F3 to F6 range such that no resulting final sample would be transposed up more than one semi-tone and/or down two semi-tones. Transposing up or down more than these limits negatively affect sound quality (obvious sample speed-up/slow-down). The entire sampling session is converted to WAV format and then transferred to a PC where Sony Sound Forge Audio Studio is used to review the sung syllables and to select the best one at each pitch. Each selected syllable is saved in its own WAV file. The selected syllables are tuned with Celemony Melodyne. The tuned syllables are the formants for the vocoding phase.

Sony Sound Forge is a solid audio editor. I can work fast in Sound Forge and its “Copy new” function is ideal for cherry picking a recording session. In a few cases, I had to amplify a sample to compensate for low level. When singing across such a wide range of pitches, one needs to rely on electronics/software for amplification in order to avoid voice strain! For tuning, I used the trial version of Celemony Melodyne Single Track which installed with Sonar X3. Although the procedure to enable the trial period was wonky, Melodyne is a great tool and I will very likely buy a copy.

In the second major production step, the formant syllables are sent to the PSR-S950 vocoder and vocoded syllables are recorded on the Micro-BR. The S950 vocoder is not a true synth vocoder. (The Motif/MOX and Tyros vocoders are true “synth” vocoders.) The S950 vocoder is part of its vocal harmony proceesor. Its “VocoderMONO” mode is designed to let (untrained) voices sing into a microphone and impose the formants onto a rather natural sounding, pitch accurate synthetic voice sound.

My early investigation found that the PSR-S950 vocoder needs clean formants that are near the desired final pitch. By clean, I mean formants that do not overdrive the vocoder input and are relatively free of the (un)natural gurgles and what not in the sounds made by the human vocal system. (Well, my vocal system anyway.) The first major step in the overall process let me select the cleanest formants. However, attempts to sing outside one’s natural vocal range introduce gurgles and rasps at the low end and off-pitch histrionics and screeches at the high end. The first major process step choses the cleanest formants and tunes them to the desired pitches.

I loaded the formant samples into a Roland RD-300GX piano as an Audio Key set. Each formant sample is assigned to a particular key and is played by the RD-300GX when the key is struck. Basically, this arrangement gives me a simple one-shot playback engine. The output of the RD-300GX is connected to the microphone/line input of the PSR-S950 in order to drive the vocoder. The mono output of the PSR-S950 is connected to the Micro-BR.

Once everything is connected and levels are set, a little trial and error is needed to find the best formant at each desired vocoder pitch. Think of this as a dry rehearsal for the final recording. Frequently, the formant at the same desired vocoder pitch is the best choice for the vocoded sample. However, sometimes one of the nearby formants is better or produces a more consistent timbre or articulation across the multi-sample. This involves a lot of critical listening and A/B comparison, producing a list of formant and pitch pairs.

Then, it’s time to hit RECORD and capture the vocoded samples by playing the desired pitch on the S950 keyboard and playing the corresponding desired formant on the RD-300GX. Once again, the recording session is converted to WAV format, is transferred to the PC, and is separated into individual WAV files.

At this point, the DOT, BOP and DOW one shot samples are pretty much complete. The DOO samples need to be looped. For some zany reason, Sony Sound Forge Audio Studio saves loop points in Acid METADATA within a WAV file. The Yamaha voice editor does not pick up this information. After searching the Web, I discovered that loop info within a WAV file is not really standardized. Given that the target tool is from Yamaha, I decided to use Yamaha’s Tiny Wave Editor (TWE) to loop the vocoded DOW samples. This worked out pretty well as TWE’s crossfade looping eliminated some bad thumps without introducing artifacts. A lot of trial and error was still involved in choosing the loop points, however. TWE can be found for free on the Web, by the way.

The final production step is to bring all of the vocoded samples into the Yamaha Expansion Voice Editor (EVE) and produce the final voice as part of an S750/S950 expansion pack. I made five voice patches:

  1. DooLoops: DOO syllables over MIDI velocities 1 to 127
  2. GetLayeredUp: All syllables, velocity-switched
  3. DatStuff: DOT syllables over MIDI velocities 1 to 127
  4. BopOnPop: BOP syllables over MIDI velocities 1 to 127
  5. Dow2008: DOW syllables over MIDI velocities 1 to 127

The multi-samples are most easily tested and normalized individually. Plus, the DOO loops and other syllables are musically useful by themselves without velocity switching. I built the GetLayeredUp patch after testing the individual multi-samples and normalizing the volumes of the individual samples within. Choosing the patch names was really fun! (Apologizes to George Clinton.)

The Yamaha Expansion Voice Editor is a trial version for which the trial period was, ahem, adjusted. Yamaha needs to just face facts and release an official version of EVE. Zillions of S750/S950 people are already using EVE and if Yamaha is somehow trying to protect its expansion pack franchise, well, that train done left the station a looooooooooong time ago. At this point, an official EVE would enhance the PSR product ecosystem and sales.

EVE does not implement velocity levels/switching. I used V. Muller’s version of the OLE Toy binary editor to set the element velocity ranges in the GetLayeredUp patch. Thank you, V. A huge amount of effort went into the analysis of YEP files and Python coding and he deserves all of the credit.

Thanks to vocoding, the final samples have a consistent sound. They are a little bit plain Jane by themselves, however. I gave each patch a little bit of reverb (reverb send level 20). I also added the “Ensemble Detune 2” DSP effect (send level 64). This is a truly spiffy effect — a chorus without modulation that gives the impression of an ensemble of slightly detuned voices. It is exactly the kind of gloss that the scat voices need.

Although the velocity ranges in GetLayeredUp are reasonable, users should still expect to tweak the keyboard velocity sensitivity and touch response to their personal needs. For example, I need to play GetLayeredUp on the softest touch setting. Your mileage will definitely vary!

Please stay tuned for the initial release of the expansion pack and multi-samples.

Vocaloid is not just for anime!

As I mentioned in my last post, I’m developing a new sample-based voice for the Yamaha PSR-S750/S950 arranger workstations. Roland is famous for its “jazz scat” voice which uses velocity-switching to trigger syllables like DOO, DAT, BOP and DOW at pitch. This synth voice is good for a cappella-like arrangements (think “Take 6”) or free melody lines. It’s a real boon for those of us with weak natural voices and technique.

The Roland scat voice incorporates samples from the Spectrasonics Vocal Planet library produced by Eric Persing and Roby Duke. Although these are great sounds/samples, I want to distribute both the workstation voice (as an expansion pack) and the samples within. I intend to make my work available under a Creative Commons attribution license. Thus, I want and need to produce fully original samples in order to avoid copyright and licensing issues.

The quest

These goals and desires launched a month-long quest for suitable “scat” samples. I decided to base the scat voice on the four syllables DOO, DOT, BOP and DOW where the DOOs are looped and the other syllables are one-shots. The DOOs are triggered at relatively low velocity and provide a pad-like bed while the DOTs, BOPs and DOWs provide short staccato accents/melody. The voice implementation requires a set of multi-samples for each syllable where the multi-samples are spread across the natural range of the human voice (F3 to F6 where C5 is middle C). has a few individual sounds, but nothing in the way of multi-samples across a range of pitches. I next decided to try sampling my own voice. A few tentative attempts left me highly discouraged! I’m a baritone with a relatively small range — definitely not F3 to F6! Plus, I lack training and my technique is not particularly good.

I then began to experiment with vocoding. I was hoping to achieve loopable, pitch-accurate samples by using my voice as a formant and imposing my voice on a pitch accurate synth sound (the carrier). I experimented with the vocoders in the PSR-S950 and the Yamaha MOX6 workstation. The MOX6 vocoder is great at producing dance-floor sounds, but not so good at producing more natural vocal sounds suitable for jazz.

Not to be too cagey, I eventually found good use for the S950 vocoder and will describe this process in a separate post. Before I went in that direction, however, I discovered and tried Yamaha’s Vocaloid.


Here is how Yamaha describes Vocaloid.

Vocaloid is a technology for singing voice synthesis developed by Yamaha, and the name of this software application. The software allows users to input melody and lyrics in order to synthesize singing. In other words, with this technology, singing can be produced without a singer. Singing voice synthesis is produced by using fragments of voices recorded from actual singers, called the Singer Library.

To a user, Vocaloid consists of two parts: the Vocaloid editor and one or more libraries. Generally, Yamaha does not provide the libraries and prefers to license the Vocaloid technology to third parties (like Zero-G) who develop libraries using their own artists.

Vocaloid has an active and enthusiastic on-line community among anime enthusiasts. There are Japanese and English singer libraries for various anime characters or personas. These singers are not appropriate for jazz! Fortunately, there are a few singer libraries for pop and classical vocals.


Vocaloid is not inexpensive. The full Vocaloid version 3 editor is about $160USD and individual Vocaloid 3 singer libraries are $150USD. Thus, it’s hard to take a casual drive by the latest Vocaloid technology and give it a try. Vocaloid 4 has just been announced along with Cyber Diva. Pricing, unfortunately, has not budged.

Luckily, Zero-G has a fire sale on a few individual Vocaloid 2 libraries which include the version 2 editor. I bought the Zero-G Tonio library for $50USD. This is a much smaller amount to gamble in order to get a taste.

Tonio is an opera singer. The Tonio demo is very good (it’s opera!) and after messing with Vocaloid and Tonio, someone sank a lot of work into that demo! You can get very nice results from Vocaloid if you are willing to spend countless hours tweezing a performance. I recommend the on-line Vocaloid reviews at Sound on Sound Magazine. The reviews are right on the money and provide useful information to help get you started with Vocaloid. (SOS is great that way.)

To make a long story very short, you edit the vocal performance in the editor by entering lyrics into a piano roll editor. You then change the attack, vibrato and other aspects of the vocal performance. These tweaks are essential for getting a good result.

Ultimately, Tonio is an opera singer and his vocal characteristics are a distinct part of the vocal samples that underlie the singer library. There ain’t no way to turn this nice Italian boy out and make him sing pop! He isn’t Bruno Mars. Please keep this in mind if you decide to try Vocaloid in a project of your own. Make sure that the voice library is a simpatico match with the target genre/style. This is why I moved on from Tonio and Vocaloid for the scat voice project.

The technology

Yamaha has invested heavily in the Vocaloid technology and have filed many patents. They are conducting joint research with The Music Technology Group (MTG) of the Universitat Pompeu Fabra in Barcelona. The MTG, by the way, are the people behind the web site.

Vocaloid does a lot of intense digital signal processing (DSP). It modifies and concatenates sound in the frequency domain. It performs a Fast Fourier Transform (FFT) to convert from the time domain to the frequency domain, modifies the spectral characteristics of the sound, and then performs an inverse FFT to return to the time domain. This is too much computation to perform in real-time. Thus, there is always a delay while Vocaloid renders a performance before playback.

Yamaha protects its intellectual property (IP) through patents and rarely publishes results in the scientific literature. Vocaloid is an exception, probably due to the partnership with MTG. Here is a short list of a few papers on Vocaloid and its technology.

  • Singing synthesis as a new musical instrument, Hideki Kenmochi, IEEE International Conference on Acoustics, Speech and Signal Processing, 2012 (ICASSP 2012).
  • Sample-based singing voice synthesizer by spectral concatenation, Jordi Bonada and Alex Loscos, Proceedings of the Stockholm Music Acoustics Conference, August 6-9, 2003 (SMAC 03).
  • VOCALOID – Commercial singing synthesizer based on sample concatenation, Hideki Kenmochi and Hayato Ohshita, International Speech Communication Association (ISCA), Interspeech 2007.

You don’t need to know all of this to use Vocaloid, but it’s good to know that there is cutting edge science behind the product.

I strongly recommend the developer interview with Michael Wilson which is published at the Vocaloid US web site. The interview gives insight into the incredible amount of work and detail behind the development of the latest library, Cyber Diva. This interview is extremely informative. Thanks, Michael. Articles such as this one bridge the gap between vacuous press releases and scientific papers giving everyone a greater appreciation for the technology behind a product.

It is also the best case to be made against software piracy. Innovation, research and development is fueled by money. Cheat developers out of their just payment only if you wish to kill off future innovation!

The Vocaloid technology reminds me a little bit of Super Articulation 2 (SArt2) on Tyros. SArt2 concatenates tones together to product realistic articulations such as legato and glissando. SArt2 works in the time domain and computes in real time although latency remains a very practical concern. (There are patents.) Perhaps someday when sufficient parallel processing resources are inexpensive, there will be an SArt3 that computes in the frequency domain.

A dive into some old Roland gear

I haven’t taken a deep dive into any gear lately and my ears need a break! So, here goes.

The Web is a wonderful source of distractions. After discovering the new Roland Sound Canvas app for iPad, I was searching for some information about the later model Sound Canvas modules and came across the Roland Service Notes for the XP-80 workstation, the XV-1010 module, and the SK-88Pro Sound Canvas. I still have these instruments in my sonic arsenal although they don’t see as much day-to-day use anymore. The orchestral sounds, in particular, hold their own today and I occasionally play the XP-60 at my church gig. Some of the sounds like the Voice expansion board (taken from Vocal Planet) are unique and stellar.

Roland have their “Service Notes” and Yamaha have their “Service Manuals.” To each his own name. However, keep these terms in mind when searching the Web for documentation on the innards.

As you might expect, there are similarities to Yamaha’s approach and some differences. These instruments are embedded computer systems for Pete’s sake — they just happen to make joyful noise. These instruments were developed in the late ’90s and early 2000s.

Roland — like Yamaha — drew from the Hitachi (now Renesas) processor families. The XP and XV use the more powerful SH-1 series CPUs while the SK uses the less powerful H8/510.

All of instruments have a custom tone generation chip. Roland’s naming convention for the tone generator ICs does not suggest how these chips fit into a family although the “XP” mnemonic is used throughout. The tone generators each have a dedicated RAM for effects processing similar to the Yamaha SWP51 series. The tone generators have a dedicated wave memory interface. The Yamaha SWP51 supports two 16-bit channels while the Roland tone generators have a single 16-bit data channel. I’m just know starting to appreciate how tone generation is a bandwidth-sensitive application and the single vs. double channel difference may be functionally significant. Certainly, the Yamaha approach requires many more pins, but it supports higher bandwidth. More polyphony? More voice elements? Hmmm.

Since these are older instruments, the wave memory size is not about to knock you out. However, this is an important reminder that wave memory size isn’t everything. Roland developed and programmed very good sound sets for these instruments. The wave memory may seem miniscule in the era of multi-gigabyte sample libraries, but it’s what hits the ears which matters the most.

Here are the details — straight, no chaser. Speaking of which, I’m knee deep in the development of a scat vocal voice for the PSR-S950. This project has taken a month so far and I’m only about 90% finished. Stay tuned.

Roland XP-80

Following is a list of the most relevant bits of sand in the Roland XP-80 workstation:

IC5     XP         Unidentified      Custom tone generator

IC12    SH CPU     HD6437034SC66F    Hitachi SH7034 Superh RISC engine
IC36    WAVEROM B  LHMN5PNA          4MByte WAVE ROM for tone generator
IC37    WAVEROM A  LHMN5PN9          4MByte WAVE ROM for tone generator
IC10    MASK ROM   TC5316200CF       MASK ROM on system bus (data+code)

IC17    SRAM       HM62864LFP-7SLZ   SRAM on CPU system bus
IC19    SRAM       HY6264ALJ-70TE2   DRAM for LCD controller
IC14,15 DRAM       LH64256BK-70      EFX DRAM for tone generator
IC21,22 DRAM       HM514800CJ-80     DRAM on CPU system bus

IC39,41 DAC        PCM69AU-1/T2      24-bit DAC

The SH7034 is a 32-bit RISC engine (SH-1 architecture) with peripheral interface logic integrated onto the chip. The SH7034 chip also has a 64KByte ROM/EPROM and a 4KByte RAM. The CPU has 32-bit datapaths, sixteen 32-bit general registers and a five stage pipeline. It has a multiply and accumulate (MAC) unit.

The XP tone generator (TG) is not explicitly identified in the parts list. It is merely labeled “XP” in the circuit schematic. The tone generator interface to the WAVE ROM has 23 address bits and is 16-bit data parallel. The WAVE ROMs provide data as 16-bit words. Each ROM has 21 address bits and the 16-bit data bus is shared between all ROMs. Each ROM is 4MBytes for 8MBytes total WAVE ROM. This checks with Roland marketing literature.

The Burr Brown PCM69AU is a dual 18-bit DAC capable of 16x oversampling. This part was also manufactured by Texas Instruments.

Roland XV-1010

The XV-1010 is a one expansion slot sound module in the JV/XV family. The XV-1010 has the basic XP sound set plus the “Session” expansion sounds — essentially a built-in SRJV expansion board. The XV-1010 does not have much of a user interface and not much is required in terms of processor power to support it.

Here is a table of the main ingredients in the XV-1010:

IC1    CPU            HD6437016F28
IC2    SRAM           TC551001CF-70L      RAM on CPU system bus
IC3    XP6            RA09-002            Custom tone generator
IC4    DRAM           MN414260DSJ-06T1    DRAM on CPU system bus
IC5    DRAM           MN414260DSJ-06T1    Tone generator effect DRAM
IC6    WAVE ROM A     LHMN0PVW            8MByte WAVE ROM
IC7    WAVE ROM B     LHMN0PU5            8MByte WAVE ROM
IC14   FLASH MEMORY   LH28F160S5T-L70     FLASH memory on CPU system bus
IC26   DAC            AK4324-VF-E2        DAC driven by tone generator

The Hitachi SH7016 is another member of the SH-1 RISC processor family. Peripheral interface logic is integrated onto the chip making it ideal for embedded applications (like a synthesizer). The CPU has 32-bit datapaths, sixteen 32-bit general registers, and a five stage pipeline. The CPU has a multiply-accumulate (MAC) unit. The SH7016 has 64 KBytes of mask ROM and 2 KBytes of RAM (when the 1KB cache is enabled).

The tone generator IC wave memory interface has 23 address bits and a 16-bit data bus. Each WAVE ROM has 22 address bits and they share the 16-bit data bus. The basic XP sound set is stored in an 8MByte ROM and the Session sound set is store in an 8MByte ROM for a total of 16MBytes of wave memory.

The EXP-B SLOT has 21 address bits and an 8-bit data bus. The narrow data bus is probably a legacy artifact from the earlier JV-series instruments.

The Asahi Kasei (AKM) AK4324 is a 1-bit, 128x oversampling stereo DAC with
a maximum 96KHz sample rate. It includes a 24-bit digital filter.

Roland SK-88Pro

The SK-88Pro is an SC-88 Sound Canvas built into a 3 octave MIDI controller with the Roland two-way pitch bend/modulation stick. The SK has two MIDI IN ports with 16 channels each. The controller is built like a tank and is almost as rugged as made-of-metal Roland workstations. I would not trade this thing for any of the toy-like controllers on the market today!

Here is a table listing the important integrated circuits:

IC1     CPU              HD6415108F          H8/510 microcontroller
IC3     XP               RA01-005            Custom tone generator
IC4     CUSTOM DSP       MB87837PF-G-BND     Unique to the Sound Canvas?

IC8     PROGRAM ROM      LH538U29            Main program and data
IC17    WAVE ROM         UPD23C32000AGX-314  4MByte WAVE memory
IC13    WAVE MASK ROM    LHMN0PNM AB         4MByte WAVE memory
IC15    WAVE MASK ROM    LHMN0PNN CD         4MByte WAVE memory
IC12    DSP RAM          NN514260J-60        256K x 16-bit word memory

IC50,51 DAC              UPD63200GS-E2       DAC for each channel

IC5     SUB CPU          M38881M2-152GP      Handles MIDI input/output
IC7     SUB PROG ROM     LH5S4H0B            Program and data ROM

The Hitachi H8/510 microcontroller family uses a 16-bit CPU core. The H8 has 16-bit datapaths, eight 16-bit general registers and a maximum speed of 10MHz. The H8/510 has integrated peripheral logic and is designed for embedded applications.

The SUB CPU handles MIDI input and output nothing more, nothing less. This is different than Yamaha’s SUB CPU which does heavy lifting.

The custom DSP IC communicates with the XP tone generator. The custom DSP has a dedicated 256K x 16-bit word DRAM. The tone generator has a dedicated 256K x 16-bit word DRAM for effects processing.

The XP tone generator has three WAVE ROM ICs: IC13, IC15 and IC17. The three WAVE ROMs have 21 address bits each and share a 16-bit data bus. The tone generator drives 22 address bits; the least signficiant bit, WA0, is not connected. Address bits WA[1:22] are sent to the WAVE ROMs. The high order bit, WA22, appears to function as a (ROM) chip select. This implies 12MByte of WAVE ROM total.