I became interested in exploring what’s currently available for offline speech recognition using neural networks.
Yandex’s Alice assistant informed me about a project called Vosk:
https://alphacephei.com/vosk/
I decided to integrate this into NVDA - perhaps it will be useful to someone. I’m attaching the result to this message.
https://t.me/nvda_group/60655
First, install the main recognition engine (file VoskSTT.nvda-addon), then additionally install the recognition model for the target language.
In the “Input Gestures” dialog, under the Vosk STT subcategory, assign an available key combination to the command “Starts or stops speech recognition.”
After the first invocation of this command, you need to say something into the default microphone, then stop the recording by invoking the command again. NVDA will speak the recognized text and place it into the clipboard.
The settings dialog can be opened via NVDA menu -> Preferences -> Vosk STT Settings. There you can select the recognition model, the microphone to use, and enable/disable the sound signal at the start of recording.
The VoskStt engine has been tested with small voice models averaging in size from 30 to 300 MB.
Note from the manager of the Russian-speaking community:
Target language models can be downloaded as add-ons on the following page in the Russian add-ons catalog.
Additionally, they are available through the RuStore and Store add-ons, as well as TiendaNVDA with the added Russian server.
You can find all models by searching for “VoskSTT”.
I have compiled 34 small language models for the following available languages:
Notes on some languages:
Large language models were not compiled and placed in the add-on store due to their excessive size. You can compile them yourself following the example of the already compiled models.
Also keep in mind the following: large models use up to 16 GB of RAM during their operation. Therefore, a 64-bit version of NVDA is required to run them. 32-bit applications cannot address more than 4 GB.
All models, both large and small, in zip archives can be downloaded from the developer’s website.
https://alphacephei.com/vosk/models
If you liked this add-on and want to express gratitude to its author, you can do so by transferring any amount to the following YooMoney wallet: https://yoomoney.ru/to/410012293543375
Initial release.