DoubleTalk PC SAPI5

A native Windows SAPI5 voice engine for the RC Systems DoubleTalk PC, an early-1990s ISA text-to-speech card, running under MAME component emulation. No SAPI4, no DOS, no hardware — the card's original firmware runs on an emulated 80C188EB CPU and its audio is handed straight to SAPI5.

Works with any SAPI5 application: NVDA, JAWS, Narrator, Balabolka, Bookworm, and anything else that speaks through Windows.

What you get

• 32-bit and 64-bit engines, each registered in its own registry view, so both 32-bit and 64-bit applications see the voices.
• All eight documented card voices — Perfect Paul, Vader, Big Bob, Precise Pete, Ricochet, Biff, Skip, Robo Robert.
• Every card parameter exposed: rate, pitch, volume, articulation, expression, formant frequency, reverb and tone, plus the host-side reconstruction filter and a pitch-preserving rate boost.
• A configuration utility built for keyboard and screen-reader use, where every change is applied, spoken and saved immediately.
• Real word-boundary and bookmark events, timed from the firmware's own index markers rather than estimated from text offsets — so highlighting follows the audio exactly.

Two things worth knowing about the hardware

There are five distinct voices, not eight. The manual documents eight names, but this firmware takes the nO voice number modulo 5: Biff, Skip and Robo Robert render byte-for-byte identically to Perfect Paul, Vader and Big Bob at their defaults. All eight are still offered — they are the card's documented names, and they diverge as soon as you move any per-voice parameter. Verified by rendering the same sentence on a freshly created emulator instance per voice and comparing the PCM.

The card is English-only. There is no language-select command anywhere in the DoubleTalk command set — all 26 command letters were probed against the ROM and cross-checked against RC Systems' own command summary. The "English or Spanish" claim in RC's literature refers to the later RC8660 chip, and even there Spanish requires a user-supplied downloadable exception dictionary; this card's firmware does not implement the download commands at all.

Parameter ranges

These are the firmware's real ranges, established by sweeping each command against the ROM — not the later RC8660 datasheet's, which differ.

Command	Parameter	Range
nO	Voice	0–9, wraps mod 5
nS	Rate	0–9
nP	Pitch	0–99
nV	Volume	0–9
nA	Articulation	0–9
nE	Expression	0–9
nF	Formant frequency	0–9
nR	Reverb	0–9
nX	Tone	0–2 (bass / normal / treble)

Every value above wraps modulo its range, so 15A is 5A.

Settings and the registry

Tunable state lives in an INI file, not the registry:

• %LOCALAPPDATA%\DoubleTalkSAPI\settings.ini — your settings
• %PROGRAMDATA%\DoubleTalkSAPI\settings.ini — machine defaults, used only when a user has no file of their own

The engine re-reads the file whenever it changes, so an adjustment in the configuration utility is audible on the very next thing spoken — no restart.

The only registry keys this project writes are the three SAPI needs in order to find the voices at all:

HKLM\Software\Classes\CLSID\{0B5F5547-...}\InProcServer32
HKLM\Software\Classes\CLSID\{C54B5B3F-...}\InProcServer32
HKLM\Software\Microsoft\Speech\Voices\TokenEnums\DoubleTalk

The eight voices themselves have no registry presence — they are served from memory by the token enumerator. If HKLM is not writable the engine registers under HKCU instead, though note that SAPI only discovers voice enumerators from HKLM, so a per-user registration is useful for testing rather than for everyday use.

Logging

Both the engine and the utility log to %LOCALAPPDATA%\DoubleTalkSAPI\logs\, one file per component, architecture and process, capped at 4 MB with one rotation. The level is set in the configuration utility (LogLevel 0–4) and defaults to Normal. The installer always writes its own log; find it via %TEMP%\Setup Log*.txt.

Building

build_all.bat

Requirements: Windows 10+, Visual Studio 2022 Build Tools (C++ workload), CMake 3.15+, and — for the installer — Inno Setup 6.

The build produces both architectures, runs the engine self-test on each, and compiles the installer.

Tools

dt_render.exe renders text to a WAV through the same code path the SAPI engine uses, which is the first place to look when deciding whether a problem is in the engine or in the SAPI plumbing:

dt_render --self-test
dt_render --list
dt_render --voice 3 --reverb 7 --out pete.wav "Hello there"
dt_render --all-voices --outdir samples

dt_sapitest.exe drives the real SAPI stack — SpVoice, format negotiation, the event queue — while creating the voice token from our own enumerator, so the engine can be exercised without an elevated install:

dt_sapitest --list
dt_sapitest --events "One two three"
dt_sapitest --all --outdir sapi_samples

Note that SAPI only queues events to the voice when speaking to a real audio device; with --out the events go to the file's own stream instead, so --events reports none.

Credits

Two projects made this possible, and CREDITS.md sets out exactly what came from where.

doubletalk-pc by David Sexton — BSD-3-Clause. Everything that actually produces sound is that project's work: dtalk.dll and dtalk64.dll are its standalone emulator builds, a vendored MAME 80C188EB core plus the board simulation that runs the card's firmware. The C API this wrapper drives is its dtalk.h, and its NVDA driver was the reference for how the card's command set must be sequenced.

BestSpeech SAPI5 wrapper by Gozaltech — this project began as that wrapper. Its COM plumbing is reused directly: the IUnknown/IClassFactory template layer, the reference counting, the registry helpers, and the in-memory IEnumSpObjectTokens approach that keeps individual voices out of the registry entirely. If this is useful to you, they accept donations at https://paypal.me/gozaltech.

RC Systems (https://www.rcsys.com) made the DoubleTalk PC hardware, its firmware, its command set, and the eight voice names.

Licensing and the firmware ROM

The wrapper is BSD-3-Clause — see LICENSE.

doubletalkpc.bin, the 512 KB firmware ROM, is proprietary to RC Systems and is not covered by that licence. It is not in this repository. It is bundled inside the installer on the Releases page so the voices work out of the box; if you are RC Systems and want that removed, open an issue and it will be taken down.

Building from source instead? Supply your own dump and put it next to the built binaries (or in the install directory). The expected image is CRC32 66685631, SHA-1 bf7e78d6381c76d291ee069971873347a314ffff — the engine checks the size at load and logs a clear error if it is missing or wrong.