🛑 Final Update:

Added Nvidia Parakeet as a streaming provider with heavy customization, fixed several bugs with some options, moved some options around, added update log, added "finalize text cleanup (experimental)" which adds basic grammar to text, added interim results to AssemblyAI. Nvidia Parakeet only has two models and they are both >1GB. If you use Nvidia Parakeet, I'd recommend enabling prewarm and cache model on startup for the best performance.

⭐ New:
Major Compatibility Fixes, Performance Improvements for Google V1 Provider, and Updated Localizations.

❗Important:
The site that fetches the vosk models list (alphacephei.com) is having expired certificate issues at the time of this update, you can temporarily bypass it by proceeding to the site. I've included a notification that tells you how to do it.

Fixed broken vosk

Added Google as a provider with 3 models (on-par with Google Chrome), fixed rare but bad GC memory leaks, and other minor QOL fixes.

❗ Any bugs reported or feature requests after this version ARE NOT guaranteed as I'm focusing on other things.

New:
- Extended Language Support, you can now use ISO 639 language codes in addition to the presets.
- Toast that tells you if you mis-match the set language and language models when using vosk engine.
- Moved 'Languages' from vosk and whisper model categories to just the engine category for seamlessness.

Fixes:
- Mic would stay on due to a duplicated getUserMedia call until GC clears.
- Not being able to select a vosk model and similar input boxes due to "manage passwords" popup.
- Regression of set language based on site's request (ex: google translate).
- Other small annoyances i forgot to mention

FULL FIX:
Ultimately fixed the speechrecognition polyfill not initializing in sites (whole point of extension btw 😭)

PARTIAL FIX:
Fixed webkitSpeechRecognition ReferenceErrors by re-rewriting and re-implementing the then removed asynchronous polyfill injection method to execute faster than site scripts.

❗full fix in 1.5.5❗

Bug Fixing:
- Unwarranted Bump to v1.5.2 cus mozilla didnt listen when I said "cancel" a unfinished version
- Fixed language models not working with Vosk
- Fixed Garbage Collector not working with Vosk
- Added simpler language model names
- Fixed keybind double cancel regression
- Fixed exporting with Vosk
- Fixed AssemblyAI not using the language selected
- Fixed keybind having inverse functions
- Actual last update

Accidental use 1.5.2

🎊 v1.5 Final Release:
- Continuous Speech (Long awaited)
- Realtime Audio AssemblyAI Streaming (Server, very fast)
- Realtime Audio Vosk Streaming (On device, very fast)
- Less cramped developer options and other options in general
- Customizable grace window per site and a few extra options
- Ability to disable processing time out
- Alot More

New:
- Removed speech/word limit
- Fixed WASM being used intermittently despite having WebGPU supported
- Fixed issues with indicators on the microphone icon for the most part
- Added the ability to cancel speech (double tap) while its being processed
- Made it easier to restart you speech recording
- Google Docs/Slides now supports your custom keybind!
- Added scrollbar to overrides list
- Several New Developer Options (All can be exported)
- Customizable Mic Gain
- Customizable Silence Sensitivity
- Ability to hide warning message
- Separated debug mode and toast notifications

❗ FULL RELEASE

1. fixed google docs glitch for good
2. fixed stale session
3. fixed not being able to say single words
4. rewrote some parts of background.js
5. Made base-multilingual the default model
6. moved distil-medium out of experimental
7. fixed compatibility on most other sites
8. a few new developer mode options (such as microphone gain!)

Note: You may see the word Provider/Engine. They are interchangeable.

Fixes:
Inability to type in Google Docs or similar canvas-based sites within the google domain.

Known Issues:
1. Using Google Docs with the Speech to Text Keybind currently does not work (Mitigation: You have to use google's built in voice-typing for it to work). This will be fixed in the next build.
2. Stale session (or constant unintelligible/red mic icon) sometimes after a model is switched. This can be mitigated by reloading the site and will also be fixed in the next build.

✅ Guaranteed: Future Roadmap
1. Merge the keybind that opens google docs' built in voice-typing with the extension's keybind.
1.1. I might just make an option to merge the keyboard yourself for any website so It's not manually done/updated by me each time, but I'll add the major sites like google docs/microsoft word keybind merges as defaults in the extension.
2. Cancel/stop all voice session ids if the model is changed. This should fix red mic/stale session issues without having to reload the site.
3. Add an option to disable the extension for a specific site.
4. Move distil-medium out of experimental category since its decent now.
5. Make base-multilingual the default model (tiny multilingual sucks at english transcription and since the majority using this will be english speaking they may think its not working)
6. All of this should come at the same time with the full release of v1.0

❓ Not Guaranteed: Future Roadmap
1. Adding continuous voice speech recognition (prob unlikely since it'll be very hacky and annoying to deal with)
1.1. Currently, you can just click the mic again after speaking and the extension already records your audio until you finish talking anyways so I don't see the point. Would probably only work out well with the cloud api.
2. Fix the extension not detecting voice if user only says a singular word and making it work properly in Mozilla's Webspeech color test
3. If none of the above comes out for the v1.0 full release, it probably wont come out at all and was likely cancelled to maintain the current stability of the extension.

Added the ability export and import settings.
Added the ability to edit sites by clicking on them and clearing inputs for the sites lists.
Added auto-updating of site lists on options page.
Added a drop down for sites lists on options page.

TLDR: less memory bloat, better per-tab status, fewer stuck transcriptions, a hotkey, and a more reliable “Enter/submit” after dictation.

Enhancements
- Dual engine support: choose on-device Whisper or cloud AssemblyAI; store API key; optional prefetch/cache of default model.
- Adaptive VAD and stop logic: smarter noise floor, early “no speech” stop, optional 5s hard cap toggle, grace-window toggle.
- Per-site overrides include engine selection (plus model/language/timeout); popup auto-hides model when cloud engine is chosen and auto-saves changes.
- Options UI: engine card, dev toggles (hide model sections, hide favicons, hard cap, grace, cache model), “Save all” + factory reset; options auto-unload when idle.
- Action icon UX: download/cache/done/cancel badges; cached badge on reuse; brief held error state.

Fixes
- Better cancel/stale-session handling and tighter processing timeout to avoid hangs.
- Stricter handling of silent/too-short/pathological audio as “no audio/unintelligible,” reducing bad results.
- Processing fallback timer and GC/reload behavior to prevent long-running or stuck states.
- Basically its less likely to break....

Updated Garbage Collector (GC):
Idle memory usage is now at ~4MB (after being cleared see below) originally from 1GB
The extension clears cache after 30 seconds of idle and after closing the tab or navigating to a different site.