2026.9.2

Highlights

• Choose the translation tone. HY-MT2 translations now offer Standard, Formal, and Casual styles so you can set the register that fits the note.
• Add your own style guidance. Custom style instructions provide a free-form field, capped at 500 characters, for preferences such as regional wording or a specific audience.
• Review before applying. Style changes are saved for subsequent translation jobs, and the preview makes you retry when a completed result was produced with an older style.

Compatibility

• Style controls apply to HY-MT2 translations only. Existing translation settings remain valid and default to Standard.

2026.9.1

2026.9.1

Highlights

• Translate more languages on demand. Firefox translation languages are now available from the translation picker without bundling every translation pack in the plugin. Install only the language pair you need, then translate locally after setup.

Improvements

• Choose a separate read-aloud language. Speech Kit can now speak text in a language that is independent from your dictation language, and applies a language change cleanly to an active read-aloud session.

Fixes

• Keep maintenance work current. This release includes dependency security updates and general performance and reliability improvements across the plugin and its release tooling.

2026.8.7

2026.8.7

Highlights

• Choose a higher-capacity local translation model. Tencent HY-MT 2 brings 1.8B and 7B on-device translation models to Speech Kit, each covering 38 languages. The 1.8B model is the practical choice for most computers; the 7B option is available for people who want a larger model and have room for its 4.62 GB download.

Improvements

• Open your model folder in one click. Manage models now includes Open model folder, so you can quickly inspect or remove retired downloads after an upgrade. It opens in your operating system's normal file manager.

Fixes

• Explain a missing dictation model immediately. Starting dictation without a selected model now shows a clear notice instead of silently appearing to do nothing.

2026.8.6

2026.8.6

Highlights

• Follow along while Speech Kit reads. Enable spoken-text highlighting in Read aloud settings to add a subtle underline to the sentence currently being spoken—without changing your note or selection.

• Listen to a translation before applying it. Complete translation previews can now be read aloud in the target language, so you can review how the result sounds before replacing, inserting, or copying it.

Improvements

• Use the translation models you already installed. Speech Kit now selects a compatible installed translation style for your chosen language pair, keeps unavailable styles disabled, and clearly explains when a different model is needed.

2026.8.5

2026.8.5

Highlights

• Translate naturally across 38 languages. The optional Tencent HY-MT engine adds local, offline translation between any two of its supported languages. Install its approximately 1.06 GiB model directly from Tencent in Manage Models, then choose Natural in the translation preview.
• Edit translations before they touch your note. Once translation finishes, revise the preview and use your reviewed wording with Replace, Insert below, or Copy. Loading, canceled, failed, and partial results remain protected from accidental note changes.

Improvements

• Choose speed or prose style. Fast & literal keeps the compact Firefox Translations workflow for its 14 English-anchored directions. Natural is more paraphrastic and automatically handles language pairs Fast does not support.
• Recover cleanly from long translations. Natural translation now has immediate cancellation, clearer errors, elapsed-time and progress feedback, and a safe retry path without leaving a hidden job running.

Compatibility

• Review Tencent's terms before installing Natural translation. Speech Kit does not redistribute the model and requires confirmation that you are outside the European Union, United Kingdom, and South Korea. The installer links to the exact pinned Tencent HY-MT model license.

Known Limitations

• Treat Natural output as a draft. Natural is a style choice, not a promise of higher accuracy; it may paraphrase or change meaning. Speech Kit always requires preview before replacing or inserting note text.

2026.8.4

2026.8.4

Improvements

• Keep your place while the sidebar refreshes. Speech Kit now preserves the sidebar’s scroll position when its settings update.

2026.8.3

2026.8.3

Highlights

• Read aloud from the cursor. Bind the new Read aloud from cursor command to listen from your current cursor position through the end of the note. A selected passage still takes precedence, while the existing Read aloud command keeps its whole-note default.
• Copy your last dictated utterance. Use the new Copy last utterance command to place the most recently finalized speech on the clipboard without reinserting it into the active note.

Improvements

• Navigate model tasks by keyboard. The Dictation, Read aloud, and Translation tabs now support arrow-key and Home/End navigation.

Fixes

• Keep raw-transcript recovery available after renaming a note. Recovery continues to work when the same open note is renamed.
• Avoid cleanup when its target range disappears. Speech Kit now stops batch cleanup rather than sending text to a configured provider when it can no longer safely apply the result.

2026.8.2

2026.8.2

Highlights

• Use Speech Kit in Croatian. Dictate with Whisper or Nemotron, read notes aloud with Supertonic, and use the complete Croatian interface translation.
• Dictate in Serbian. Transcribe Serbian speech locally with Whisper; unsupported speech features continue to fall back safely instead of claiming broader language coverage.

2026.8.1

2026.8.1

Highlights

• Use local or self-hosted language models for refinement. Add OpenAI-compatible providers such as LM Studio, configure them separately from built-in providers, and route short or long transcripts to the provider you prefer.
• Turn Always-on dictation into a voice clipboard. Enable automatic copying to place each accepted final utterance on the system clipboard immediately. When batch refinement successfully replaces the transcript, its cleaned result replaces the raw clipboard text too.

Internal

• Releases start sooner. The release pipeline no longer waits for a duplicate Windows CUDA pre-build before starting the authoritative cross-platform build.

2026.8.0

2026.8.0

Fixes

• Settings work correctly in Obsidian 1.13. Speech Kit's full settings page now appears instead of an empty panel, while remaining compatible with supported older Obsidian versions.
• Settings interactions are more reliable. Microphone selection opens at the correct width, keyboard focus is preserved during refreshes, and device and hotkey controls use the correct window when Settings is opened separately.