Eloquence synthesizer add-on for NVDA with full 64-bit support.
The Eloquence DLL is 32-bit only. This add-on launches the Eloquence Host
Process (eloquence_host32/eloquence_host32.exe) to load the Eloquence Engine and stream audio
back to 64-bit NVDA. The integration is transparent — no additional Python
installation or manual steps are required.
For development scenarios where the prebuilt Eloquence Host Process executable is unavailable,
the ELOQUENCE_HOST_COMMAND environment variable can be set to the command that
launches a compatible 32-bit Python interpreter with host_eloquence32.py.
The Audio contour setting in NVDA's Speech settings shapes Eloquence's audio after it is synthesised, using Eloquence Contours. It applies to both engines alike. The contours are tied to the sample rate the engine runs at:
| Contour | 11025 Hz | 16000 Hz |
|---|---|---|
| Raw (default) | unchanged | unchanged |
| Presence | Classic Presence, played at 22050 Hz | native Presence |
| Presence treble | hozosch v19.1 treble, played at 22050 Hz | — |
| Smooth | — | native Smooth |
At any other sample rate, or with a contour that has no version for the rate in use, audio plays raw.
Eloquence can read words differently using pronunciation dictionaries. You choose which ones each language uses in NVDA's settings, under Eloquence, with the Manage dictionaries... button.
In that dialog you pick a language and get its list of dictionaries, each with a checkbox:
- Dictionaries from an installed add-on, such as a dictionary editor that ships a community set.
- Downloaded dictionaries. The dialog lists the GitHub repositories to download from. Two community projects are listed to begin with, and you can add others with Add source.... Update Dictionaries downloads them all. Downloading never switches anything on.
- Your own files, added with Add file.... The add-on keeps its own copy, so you can move or delete the original.
- Personal entries from a dictionary editor, when you have any. They are always last.
Where two dictionaries have the same word, the lower one in the list wins. Use Move up and Move down to change the order. Personal entries win over everything.
British English, Latin American Spanish, Canadian French and Chinese have one more checkbox, such as Also use American English dictionaries. It is on to begin with, and the language's own dictionaries are applied on top.
A new installation has nothing switched on. Changes take effect when you choose OK. An entry saved in a dictionary editor is picked up within a couple of seconds while Eloquence is speaking, with no restart.
If a dictionary has entries that break the engine's rules, they are still used, because most of them work in practice. The dialog tells you how many there are and lists them in a field you can read and copy, so you can fix the ones that do not take effect.
Older versions kept dictionary files inside the add-on's own folder and merged downloads into them. When you update:
- Those files are copied once to
eciDictionaries\legacy\eloquencein your NVDA configuration folder, as a backup that is never changed afterwards. - They appear in the dialog as your own files, marked as imported. The ones Eloquence was using are switched on, so speech stays as it was.
- To start clean, switch the imported files off and choose the dictionaries you want.
Two things can sound different afterwards. Entries after a blank line in an old file are now used, where the engine used to stop reading at the blank line. And a dialect now gets its main language's entries as well as its own, where it used to get only its own file when it had one.
Everything is stored in your NVDA configuration folder, in eloquenceDictionaries
and eciDictionaries, so it survives add-on updates. To use dictionaries on the
logon screen, copy your settings to it as usual with NVDA's Use currently saved
settings during sign-in button.
When the Mandarin Chinese voice is selected, Text Preprocessing applies Script Conversion before text is sent to the Eloquence Engine. Traditional Chinese text is read via Traditional→Simplified conversion with the Mandarin Chinese voice.
This is not zh-TW support, a Traditional Chinese voice, or Cantonese support.
The add-on's Chinese Voice Identity still advertises only zh-CN.
Known limitations:
- Hong Kong (
zh-HK) users get Mandarin readings, not Cantonese. - Colloquial written-Cantonese characters, such as
嘅,哋, and咗, are unpronounceable. - A zh-TW-localized NVDA install does not auto-select the Chinese voice on first run. Users need to pick the Chinese voice once manually.
NVDA does not copy *.exe files to its Secure Screen configuration for
security reasons, so the Eloquence Host Process is missing after you click
"Use currently saved settings during sign-in" in NVDA's General settings.
The easiest way to fix this is the built-in button in the add-on:
- Open NVDA Settings > Eloquence.
- Click "Copy Helper to System Config (for Logon Screen)".
- Accept the UAC elevation prompt.
Eloquence should now load on Secure Screens. You only need to do this once per add-on update.
If you upgraded from v16 (or earlier) to v17+ and NVDA reports "Could not load the synthesizer" when you select Eloquence, the NVDA log most likely shows:
AttributeError: module 'synthDrivers._ipc' has no attribute 'create_listener'
This is caused by one or more of:
- Stale Python bytecode (
__pycache__) left over from the previous version. - A half-finished NVDA upgrade leaving an
Eloquence.deletefolder alongside the new install. - The IBMTTS add-on also being installed — running both at the same time is not supported.
To recover, do a clean reinstall:
- In NVDA, open Tools → Manage Add-ons, disable Eloquence, and restart NVDA so the disable takes effect.
- In File Explorer, open
%APPDATA%\nvda\addons\and delete the entireEloquencefolder. While you're there, delete any sibling folders whose names end in.delete. - If the IBMTTS add-on is installed, disable or remove it as well.
- Restart NVDA, then install the latest Eloquence release fresh.
- As a last resort, back up
%APPDATA%\nvdaand remove it to start with a clean NVDA config.
See issue #101 for the background.
- Python Install Manager (
.msix) - uv
- 32-bit Python 3.13:
py install 3.13-32 - Rust with the
x86_64-pc-windows-msvctarget, for Eloquence Contours
git submodule init && git submodule update # fetch submodules (Eloquence Contours)
python fetch_eci.py # one-time: download proprietary ECI.DLL + voice data
build_host.cmd # compile Eloquence Host Process (only needed if host_eloquence32.py changes)
build_contours.cmd # build Eloquence Contours (add --latest to take upstream master)
scons.bat # package everything into the .nvda-addon fileNote: scons.bat validates that proprietary files, the Eloquence Host
Process executable and the contour DLL exist, but does not fetch or build them —
the earlier steps must be done first.
The eloquence-contours submodule is pinned, and published releases build
that commit. Every other CI build runs build_contours.cmd --latest against
upstream master, and Dependabot opens a pull request to move the pin whenever
upstream changes.
runlint.bat # run Ruff using the locked uv environment
runpytest.bat # run pytest using the locked uv environmentTooling dependencies are pinned in pyproject.toml and uv.lock, following
NVDA's current dependency-group pattern. The Eloquence Host Process build uses a
separate .venv32 environment so PyInstaller can run under 32-bit Python
without replacing the normal development .venv.