Version 1.0.0.
Local multilingual speech with separately downloaded models, optional NVIDIA acceleration, NVDA audio playback, speech controls and signed GitHub updates.
After installation, NVDA offers to download the speech model. The download is approximately 1.4 GB and contains all 38 voices. Allow at least 1.5 GB of free disk space. If you decline, open NVDA Settings, LFM3 Audio, and choose Download model.
Downloads continue when Settings is closed. Reopen the panel to inspect the current file, progress, speed or any error. Cancel download removes incomplete files. Quitting NVDA or a network failure preserves partial downloads for resuming; NVDA's normal exit behaviour is unchanged. A server that ignores resume requests causes that file to restart safely. Abandoned partial downloads expire after seven days, checked when NVDA next starts. Failed downloads are not installed; you can retry. Both model files are verified before installation. The model is stored in the NVDA configuration folder, outside the add-on, so ordinary add-on updates do not download it again.
Once the download finishes, select LFM3 Audio in NVDA's Select Synthesizer dialog. Your existing synthesizer remains selected until you change it. Speech runs locally, without uploading text. Starting the engine can take several seconds, especially on first use while graphics kernels are prepared.
Requires 64-bit Windows and 64-bit NVDA 2026.0 or later. A Vulkan-capable GPU with sufficient graphics memory is recommended. GPU acceleration is enabled by default; if initialization cannot use it, the engine tries CPU synthesis instead. CPU synthesis can be considerably slower and more demanding. The Settings panel reports the active backend once LFM3 Audio is running. Changing the GPU option takes effect the next time the synthesizer loads.
With a compatible NVIDIA graphics card and driver, open NVDA Settings, LFM3 Audio, and choose Download NVIDIA accelerator. This optional CUDA runtime can improve responsiveness and rendering speed. The download is approximately 590 MB. Allow at least 1.7 GB of free space for downloading and extraction; approximately 730 MB remains installed. No driver or system-wide software is installed. Other graphics cards continue to use the bundled Vulkan backend.
Downloads continue if Settings closes. The status list shows the current file, progress, speed and errors. Cancel download removes the incomplete installation. Archives are verified before extraction and removed after successful installation. The accelerator is stored outside the add-on and retained across ordinary updates.
Reselect LFM3 Audio after installing the accelerator. With GPU acceleration enabled, the engine prefers the downloaded CUDA runtime. If CUDA cannot start, it tries the bundled Vulkan runtime, then CPU if necessary. The status list reports the backend actually in use. Turning off GPU acceleration uses CPU instead. Performance depends on the graphics card, text, selected voice and other work.
Some NVIDIA Vulkan runs can leave increased CPU activity after speech ends. The optional CUDA backend has not shown this in testing and is recommended on compatible NVIDIA systems. Switching synthesizers closes the helper.
The synth settings ring and Speech settings provide Language filter, Voice, Rate, Rate boost, Pitch and Volume. Each language has a female and a male voice. The language filter restricts the voice list; it does not translate text. NVDA's automatic language switching can choose a matching language while retaining the selected gender. Unsupported language changes retain the selected voice.
Audio uses NVDA's output device and supports pause, cancellation and playback index notifications. Rate, pitch and volume use NVDA's Sonic audio processing, not a separate add-on. Neural speech can occasionally misread text; check important names, numbers and rendered recordings.
English (United Kingdom and United States), Arabic, Danish, German, Greek, Finnish, French, Hebrew, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Swedish, Turkish and Chinese.
Model downloads connect directly to Liquid AI's model repository and its download hosts. They receive the normal network information associated with a file download, but no speech text. The add-on uses a tested revision with fixed SHA-256 checksums, not an untested moving latest version. New compatible model revisions can be adopted in future add-on updates. No recordings or speech text are saved by the synthesizer.
The optional accelerator downloads verified, version-pinned runtime archives from the llama.cpp release on GitHub and GitHub's download hosts. These requests contain no speech text.
Liquid AI created LFM3-Audio. The separately downloaded model uses the LFM Open License 1.0. Its commercial conditions include a revenue threshold; read the terms rather than assuming unrestricted commercial use. Add-on and native host code are GPL-2.0-or-later. The llama.cpp runtime is MIT-licensed; bundled third-party notices are in the licenses folder.
The optional CUDA libraries are provided by NVIDIA under the NVIDIA CUDA licence terms, not the add-on's GPL licence. Read those terms before downloading the accelerator.
Created by Andre Louis with development assistance from OpenAI Codex. Thanks to Liquid AI, the llama.cpp contributors, NV Access and Sonic's contributors. Playback and settings patterns build on Phonon for NVDA; background download state follows the Samsung Galaxy Voices manager.
NVDA Settings, LFM3 Audio offers Never, Hourly and Daily update checks and a Check for updates now button. Updates come from this add-on's GitHub releases. The signed product identity, version, download size and checksum must match. Unsigned or invalid updates are not installed. NVDA handles staging and its normal restart prompt; downloaded models and accelerator files remain separate.