Speech Recognition Polyfill (STT) door apersongithub

The next and FINAL major update which is scheduled for early February will change the default from tiny-multilingual to base-multilingual since its a better middle ground for transcription (tiny performs somewhat poorly with english), I would recommend switching to the base model ASAP. Stay tuned as the next update will include new QOL features and bug fixes.

On first install this extension will open the options page, the default model language is English but this is easily changeable. This extension allows per-site customization and a multitude of different models to decipher language. Keep in mind that this is not a complete solution and the API doesn't have full support. Speech Detection is not as instantaneous like in Google Chrome's API but the added AssemblyAI integration allows for decently fast translation. The extension icon color/indicator changes depending on the process so pin it to your menu to verify the extension is working as intended. A red mic/error icon does not necessarily mean your mic isn't working but rather the speech may have been cancelled by user input, missing cloud API key, or that it is unintelligible (usually its the latter).

Make sure you are using the correct mic and speak loud, slow, and clear otherwise your voice may not be detected or unintelligble. Change the default model to the cloud or slightly larger local ones if you experience problems with voice recognition (this may impair performance). The extension does not support continuous recording as of right now.

If you're using Duolingo or similar and are trying to do the speaking practice of the language that you are learning, set the language in the extension to the one you are learning (navigate to the site -> click extension icon -> set language then click "save for site"). This will help with speech detection since the model now knows what language you are trying to speak. (This isn't necessary for every site, one example is google translate which has the auto-language implementation) Look at the images for more help.

The extension will take ~1GB of ram on normal/cloud models and up to ~7GB if you use the biggest model (you don't need to use the biggest model lol). I've implemented decent memory management to compensate.

~~~~~~~~~~~~~~~~~~~~

❗ General Recommendations:

• 8GB of RAM is a minimum requirement since it could easily take up to a decent chuck when utilizing larger models.

• A modern CPU is recommended.

• An internet connection. Even though the model runs locally, the extension re-downloads it either when idle or after closing the tab/opening a new one that utilizes the extension (for memory preservation purposes). This is ultimately better than packaging the large models within the extension for the time being and for most models, the download speed will be near instant for the general population. We also have an option in settings to keep the default model cached without re-downloading every time. Apart from locally you can use the cloud based model which is less hardware intensive.