Using the X68000 ADPCM chip for voice recording and playback

The Sharp X68000 is best known for arcade-quality graphics, FM synthesis, and fast 68000-based performance, but its audio hardware also supports a useful form of sampled sound. The built-in ADPCM channel can store and reproduce speech, sound effects, and short musical phrases while using far less memory than raw PCM audio. With suitable software and a clean input signal, it can serve as a compact voice recorder as well as a playback device.

Working with the ADPCM chip requires a different mindset from using a modern sound card. The data is compressed, the sample rate is modest, and the available recording path depends on the particular X68000 model and accessories attached to it. Careful level setting, buffer management, and format conversion matter more than chasing high fidelity.

What the X68000 audio hardware provides

Most X68000 systems use an OKI MSM6258-family ADPCM sound source for sampled audio. It is a monaural codec intended for speech and effects rather than studio recording. The chip reduces each sample to a small adaptive code, then reconstructs an approximation during playback. This makes a voice clip practical in limited RAM or on a floppy disk, even though the result has a distinctly lo-fi character.

The hardware is separate from the system’s FM sound generators. FM music continues to be synthesized by the Yamaha sound hardware, while ADPCM data is read from memory and sent through the codec. A program can therefore combine speech or sampled effects with FM music, provided that it schedules access to the sound registers and manages the available transfer bandwidth.

Sampling rates depend on the clock configuration and software conventions used by the machine. Common X68000 ADPCM resources are associated with rates around 3.9, 5.2, 7.8, and 10 kHz. These rates are adequate for intelligible speech, but they do not capture the frequency range expected from modern voice recording. The lowest settings save space; higher settings produce clearer consonants at the cost of larger files and faster data delivery.

The recording path needs careful checking

Playback is comparatively straightforward because the computer can feed prepared ADPCM data to the codec. Recording is more complicated. A compatible microphone or line-level source must reach the system’s audio input through the correct connector and circuitry, and the input arrangement can vary between models, revisions, and modified machines. Some systems may require an external preamplifier, an appropriate cable, or a dedicated expansion device.

Before writing software, verify the electrical path with a known signal. A microphone intended for a modern computer or phone may not have the same biasing or plug arrangement as an X68000 input. A dynamic microphone, electret microphone, line output, and headphone output all produce different signal levels. Connecting an incompatible source can result in silence, severe distortion, or excessive noise.

The Nereid-X project is relevant when building or restoring an expanded X68000 workstation, since the Nereid documentation provides useful context about modern hardware additions and the practical environment around the computer. An expansion board does not automatically replace the original ADPCM input circuitry, so its documentation should be read alongside the service information for the specific X68000 model.

How ADPCM data differs from PCM

Pulse-code modulation stores a numerical amplitude value for each sample. ADPCM instead records a prediction error: the difference between the expected signal level and the actual level. The codec changes its step size as the signal varies, allowing small codes to represent quiet detail and larger steps to follow rapid changes. This is why ADPCM occupies less storage while retaining recognizable speech.

The compressed representation is not interchangeable with WAV, AIFF, or ordinary unsigned eight-bit PCM. A file converter must decode the ADPCM nibbles using the correct step table and predictor rules. Nibble order is also important. Depending on the software format, the high nibble may precede the low nibble, or the order may be reversed. A file can have the right length and still sound like harsh noise if its nibble order or header interpretation is wrong.

A practical workflow is to record or prepare ordinary PCM on a modern computer, normalize it conservatively, resample it to an X68000-friendly rate, encode it to the expected ADPCM format, and test it on real hardware. For recording directly on the X68000, the reverse process is required: capture compressed data, preserve the exact byte stream, and decode it later for editing or archival storage.

Characteristic X68000 ADPCM Raw PCM
Typical use Speech and short sound effects General-purpose audio
Storage efficiency High, because each sample uses a compact code Lower, especially at higher resolution
Common X68000 rates Approximately 3.9–10 kHz Determined by the source and file format
Audio channels Usually monaural Mono or stereo
Editing convenience Requires ADPCM decoding Directly editable in most audio tools
Best source material Clear speech with controlled levels Music, effects, and archival recordings
Main technical risk Wrong nibble order or codec parameters Incorrect sample format or byte order

Designing a simple voice recorder

A direct recorder should treat input capture as a timed data-acquisition task. At each sampling interval, the program obtains a value from the audio input, updates the ADPCM predictor and step index, and writes one encoded nibble into a buffer. Two nibbles normally share one byte, so the program must track whether the next code belongs in the high or low half of that byte.

Timing must remain stable. If samples are captured too slowly, speech becomes distorted and the pitch drops; if they are captured too quickly, playback is unnaturally fast. An interrupt-driven routine or a carefully synchronized hardware transfer is preferable to a foreground loop that can be delayed by display updates, disk access, or keyboard polling. The main program should process completed blocks while the capture routine fills another buffer.

Memory planning is equally important. A short clip may fit in conventional RAM, but a longer recording needs block storage on floppy disk, a hard disk, or an attached mass-storage device. Writing directly to a disk during every sample interval is unsafe because disk operations have unpredictable delays. Use a ring buffer or double buffer, and write only after a complete block has been collected.

A recorder also needs a clear end-of-recording marker and a way to save metadata. Store the selected sampling rate, byte count, nibble convention, and any input calibration information with the file. Without those details, a raw recording may be difficult to identify years later, especially when several utilities use similar extensions but incompatible headers.

Getting intelligible playback

Speech playback benefits from preparation before encoding. Remove long periods of silence, apply a gentle high-pass filter if the source contains rumble, and avoid aggressive normalization that pushes background hiss into the audible range. A small amount of compression can make quiet syllables easier to hear, but clipping before ADPCM conversion creates harsh artifacts that the codec cannot repair.

The playback routine should feed data at the rate expected by the selected hardware setting. It must recognize the end of the stream and stop the ADPCM channel cleanly rather than continuing to output stale buffer contents. If playback is interrupted, reset the codec state before starting another clip; otherwise, the first part of the new recording may be decoded using the previous predictor and step size.

Short phrases are a good starting point. Numbers, menu prompts, warning messages, and single sound effects reveal problems quickly because they contain clear transients and recognizable consonants. A spoken sentence can expose timing errors, excessive hiss, or clipping more reliably than a sustained tone. Test through the X68000’s normal audio output and, when possible, through a second recording device for objective comparison.

The perceived result depends on the entire analog chain. A noisy power supply, aging capacitors, poor grounding, or an unsuitable microphone can dominate the sound before the ADPCM encoder does anything. Hardware restoration is therefore part of audio preservation. Clean connectors, stable power rails, and sensible shielding can improve a recording more than changing a software encoding option.

Practical software and hardware safeguards

Keep the original capture data. Even if the compressed recording sounds acceptable, save a decoded PCM copy and document the procedure used to create it. If the source was recorded directly on the X68000, preserve the raw ADPCM file, the playback utility, and any notes about the machine model. Future tools may decode the stream more accurately, while undocumented conversions can permanently obscure the original format.

Use conservative input levels. A voice signal should have enough amplitude to overcome the system’s noise floor, but it should leave headroom for loud consonants and sudden changes in distance from the microphone. Monitoring with headphones or an oscilloscope is useful when available. If the input clips at the analog stage, lowering the digital gain afterward will only produce a quieter distorted file.

A reliable development routine can be organized around these checks:

  • Confirm the input connector, microphone requirements, and signal level before connecting valuable equipment.
  • Record a short test block at every supported sample rate and compare timing, pitch, noise, and speech clarity.
  • Verify ADPCM nibble order, predictor initialization, end-of-file handling, and codec reset behavior.
  • Use double buffering or a ring buffer so disk access never interrupts sample capture.
  • Preserve raw files, decoded audio, machine details, and conversion tools together as an archival set.

Software should also fail safely. If the capture buffer approaches overflow, stop recording and report the condition instead of silently dropping blocks. If a file header does not match the selected rate or data length, refuse playback or display a warning. These precautions are especially valuable on vintage systems, where debugging tools and storage capacity are limited.

Building a reusable X68000 sound library

Once recording and playback work reliably, the same tools can support a broader sound library. Speech prompts can be catalogued by language, speaker, sampling rate, and source machine. Effects can be tagged by duration and intended use. A small index file can record the ADPCM format, byte count, playback rate, and any loop points, making it easier for games, demos, and utility programs to reuse the material.

For preservation, retain both the historical form and a modern derivative. The original X68000 file shows how the material was used on the platform, while a decoded WAV file makes listening and analysis easier on current computers. Do not replace the original with a cleaned-up version. Noise reduction, resampling, and filtering may be useful for presentation, but they should be separate from the untouched capture.

The ADPCM channel is especially effective when paired with event-driven software. A program can trigger a short spoken message after a key press, mix sampled warnings with FM background music, or use stored phrases in an experimental interface. Because the channel is limited, thoughtful arrangement matters: reserve high-rate clips for speech that needs clarity, and use lower rates for brief effects where texture is more important than fidelity.

Working within these limits illustrates an important part of X68000 development. The machine’s audio hardware was designed around efficient storage and responsive playback, not modern recording quality. Learning its encoding rules, timing behavior, and analog constraints turns those limits into a well-defined creative target.

Try a short, documented recording first: capture a few spoken words, test each supported rate, decode the result on a modern computer, and compare it with playback from the original machine. Then preserve the raw file and the conversion notes alongside your X68000 software archive. This small experiment provides a dependable foundation for voice prompts, sampled effects, restoration work, and future ADPCM projects.

Nereid-X Expansion Board

A personally-produced LAN+USB+Memory expansion board for Sharp X68000 series computers. Multiple production runs were offered, including a final batch and a later revival reproduction run.

Power Supply Repair

X68 power supply repair and modification services were offered by the site owner, with documentation shared through diary entries spanning 2001–2006.

Server & Networking

Notes on FreeBSD administration, ISP changes, server migration, and networking topics. The site itself ran on FreeBSD with the hns diary system and Namazu search integration.

A two-ink risograph print in muted slate-blue and charcoal on off-white paper, showing a stylized desktop computer monitor beside a circuit board with soft geometric trace lines, conveying a calm retro-computing workshop atmosphere. A two-ink risograph print in deep purple and dark grey on cream stock, depicting a compact expansion card with connector ports and subtle Japanese technical annotations, evoking a hobbyist electronics bench. A two-ink risograph print in teal and charcoal on warm white paper, showing a server rack silhouette with soft network-line motifs and a small weather icon, suggesting a personal server room corner.

Get in touch

X68K.NET connects Sharp X68000 enthusiasts through community links and shared projects. Reach out with questions about the Nereid project or X68 resources.