SoftVoice SAPI 5

The 1997 SoftVoice synthesiser exposed to Windows as forty SAPI 5 voices — twenty built-in personalities, each speaking English and Spanish — usable by any speech application, in both 32-bit and 64-bit programs.

Download the installer →

SoftVoice is a formant synthesiser written by SoftVoice, Inc. and licensed to The Productivity Works, who shipped it as the speech engine inside pwWebSpeak — a self-voicing web browser from the late 1990s, and one of the first browsers designed from the start for blind users rather than retrofitted with a screen reader.

It is not a concatenative or neural voice and does not try to sound like a person. It builds speech from a model of the vocal tract, which is why it can offer twenty genuinely different characters — a giant, a martian, a fly, a choir boy — out of one engine, and why it stays intelligible at speeds where recorded-unit voices fall apart. For a lot of long-time screen reader users that trade is the point.

The engine survives today because the svWebspeak NVDA add-on keeps it running. This project takes the same engine and makes it available to everything on Windows that speaks, rather than to NVDA alone.

Each appears twice: SoftVoice Male speaks English, SoftVoice Male (Spanish) speaks Spanish. Those are the only two languages the engine has — there is no German rule set in this package. Forty voices in all, and the installer lets you choose exactly which of them to have — see below.

Hear all forty: the samples archive on the Releases page contains one file per voice plus a guided tour of each language. Or regenerate them yourself with python tools\render_samples.py.

Download SoftVoiceSAPI5_Setup_<version>.exe from Releases and run it. It needs administrator rights, because speech engines are registered for the whole machine rather than for one user.

It installs:

• both SAPI 5 interfaces, 32-bit and 64-bit, and registers them
• the engine and its host
• the voices you chose, out of the forty on offer
• the SoftVoice Speech Settings utility, with a desktop shortcut
• optionally, the diagnostic tools

Then pick a SoftVoice voice in your screen reader or in Windows speech settings. To remove it, use Apps and Features as normal.

Forty voices in the list is a lot to page through if you only ever wanted two, so the components page lets you pick them one at a time. Every voice is its own check box, grouped under English voices and Spanish voices: check a language to take all twenty of it, or open the group and choose individually. Five ready-made choices sit above the tree — pick one and adjust it, or start from Custom:

Choosing few costs nothing and gains nothing but a shorter list: the catalogue is compiled into the DLL, so no voice takes disk space of its own. What the choice really controls is which tokens the enumerator hands SAPI, and that is recorded in voices.ini beside the program — one line per voice, 1 to publish it and 0 to hide it. Change a line by hand (the folder is protected, so with an elevated editor) and the next application to ask Windows for its voices sees the change; nothing needs restarting. Deleting the file, or setting every line to 0, publishes everything again rather than leaving the machine silent.

Running the installer again is the easier way round: it opens on whatever was chosen last time, so voices can be added or dropped without uninstalling first.

The one part of the choice that is a real file is Spanish itself. Svspan32.dll is the engine's Spanish letter-to-sound rule set, and it is only laid down when you keep at least one Spanish voice — and removed again if you later drop them all.

At least one voice has to be chosen; the wizard refuses to go on otherwise, since the alternative is a speech engine Windows can find and cannot use.

The installer is built to be usable by the people most likely to want these voices: every page is standard Win32 controls that screen readers read natively, nothing steals focus, and no page auto-advances. The forty voices are check boxes in the standard components tree rather than a custom page, so each is announced with its name and its check box role as it is focused, and each carries its language in the same form the voice itself uses — Male, Male (Spanish) — so an item read on its own is unambiguous. The final page states in text exactly what happened: which interfaces registered, how many voices of each language are published, and where the logs are.

SoftVoiceConfig.exe adjusts, per voice, every parameter the engine exposes:

Everything below volume defaults to the voice's own and is sent only when deliberately chosen. That is not tidiness. Each of the twenty personalities is a complete preset, and there is no value for these that means "neutral": forcing SVSetAVBias(0), which looks exactly like a neutral, drives 14 of the 20 voices past full scale, and the engine wraps rather than saturating, so it is heard as a harsh crackle. Rate and pitch are anchored per voice for the same reason — preset rates span more than three to one, from Fast Fred at 301 to Choir Boy at 89.

A change is written to the settings file the moment it is made, and the engine re-reads that file at the top of every utterance, so it takes effect on the next thing spoken with nothing restarted, and closing the utility cannot lose it. Settings live in %APPDATA%\SoftVoice SAPI5\softvoice.ini — deliberately not the registry, so the engine can speak without reading a registry key and the utility needs no elevation to save.

A plain Win32 dialog. Every control is in the tab order, none is ever disabled, each has its own label immediately before it, and the access keys are unique.

Enumerated settings are drop-down lists rather than trackbars, because the MSAA proxy for a trackbar reports its value as a percentage of its range — a list of eight glottal sources sitting on the third would be announced as "28", a number that appears nowhere in the documentation.

tools\check_config_a11y.py verifies all of this through MSAA, the same interface NVDA uses, and fails if any focusable control is unnamed, disabled, or shares an access key.

The engine is a 32-bit 1997 DLL that creates a top-level window and is not thread-safe, so it runs in its own process (svwebspeak-host.exe) and streams PCM back over a loopback socket.

That single decision is why this package is simpler than most dual-architecture speech wrappers:

• One set of sources for both architectures. Neither the 32-bit nor the 64-bit interface loads the engine, so there is no bitness-specific code path, no COM surrogate, and no second helper to keep in step.
• No SAPI 4. The engine's own C entry points are called directly.
• Nothing about the engine in the registry. The host and SVctl32.DLL are patched to hand the SoftVoice registration number to the engine in-process, so SVRegister never opens a key. The voice catalogue is compiled in and served through a TokenEnums enumerator, so no per-voice registry entries are written either.

The only registry entries created are the two CLSIDs and the one TokenEnums key Windows requires in order to find a speech engine at all. Those must live under HKEY_LOCAL_MACHINE — SAPI reads the enumerator list from there only — which is why installing needs administrator rights.

The host renders a whole utterance in about 14 ms and hands it back in a single frame — then sits on a fixed 400 ms idle timer before reporting the utterance done, refusing to start the next one until it expires. It burns no CPU doing so; it is a timer, not work.

Waiting for that report is the obvious thing to do, and it costs 400 ms on every utterance. That lands squarely on interruption: arrow to the next item and nothing can be said until the previous utterance's timer runs out. Instead, this engine treats an utterance as finished once its audio has arrived and a brief quiet period has passed, and sends CMD_STOP, which cancels the timer. Interruption is also polled for while waiting, so a keypress is acted on within milliseconds rather than whenever the next block of audio happens to turn up.

Measured end to end through a live ISpVoice with sv_latency, against Microsoft's own voices on the same machine:

A 1997 synthesiser running in a separate process answers a keypress in about half the time Microsoft's in-process voices do. That is close to the floor: the engine itself takes 12–15 ms to render anything at all, near enough independent of how much text it is given, so roughly four fifths of what remains is the 1997 binary rather than this wrapper. Dropping to 8000 Hz saves 0.6 ms — there is no quality-for-latency trade worth making.

Both interfaces, the utility and the tools write to %LOCALAPPDATA%\SoftVoice SAPI5\Logs, one file per process, at a level chosen in the utility (off through trace). Errors are always logged whatever the level. The installer writes a full log and copies it to install.log beside the program. There is a Start menu shortcut to the log folder, and a button for it in the utility.

powershell -File tools\build_all.ps1

Builds both architectures, stages output\, runs the SAPI self-test against both DLLs, and compiles the installer into dist\.

Requirements: Visual Studio 2022 Build Tools (ATL is not needed), CMake 3.20+, Windows SDK 10, Inno Setup 6, and Python 3 for the check scripts.

Pass -Version 1.2.0 to stamp a version, -SkipInstaller to stop after staging, or -Probe to also build the accessibility probe installer — the same wizard with no payload and no elevation, for walking the pages with a screen reader without installing anything.

One measurement worth knowing before writing a test against this engine: it is not sample-exact between two identical renders. The same preset spoken twice differs in up to 7% of samples, so a byte comparison proves nothing about whether a parameter had an effect. check_params.py compares against that run-to-run noise floor instead.

• SoftVoice, Inc. and The Productivity Works for the engine, and for pwWebSpeak.
• seedy60/svWebspeak — the NVDA add-on that keeps the engine alive, and the source of the host protocol and much of the hard-won knowledge about the engine's parameters that this wrapper is built on.
• gozaltech/BstSpeech-sapi — the SAPI 5 wrapper that the COM and SAPI plumbing in src/sapi/ was adapted from.

The SAPI 5 wrapper, the configuration utility, the tools and the installer written for this project are licensed under the MIT License (see LICENSE). The files that NOTICE.md lists as adapted from other people's work (the COM and SAPI plumbing in src/sapi/, and tools/svhost.py) are not covered by it.

SVctl32.DLL, SVENG32.DLL, Svspan32.dll and svwebspeak-host.exe are proprietary 1997 SoftVoice, Inc. / Productivity Works files and are not covered by it. They are included so the engine can be built and used; no licence to redistribute them is granted or implied by this repository.