Every release.
Every detail.

What's new

• Model downloads (Qwen 0.6B and 1.7B) now try Hugging Face first, since it's much faster than ModelScope outside mainland China. On machines set to a mainland China time zone (checked locally, nothing sent anywhere), ModelScope is tried first instead, since Hugging Face isn't reachable there.

Improvements

• Downloads now recover automatically from a stuck or very slow source: if a source stays silent for 20 seconds, or averages under 200 KB/s over 15 seconds while another source is still available, the download switches sources and resumes instead of hanging or crawling forever.
• The recording prompt no longer clips the mic icon or waveform when a hint is long (e.g. the locked-recording instruction). The waveform shrinks first, then the hint text wraps onto up to two lines, breaking only at natural word/punctuation boundaries so shortcut labels like "Ctrl + Shift" stay together.

Fixes

• None user-facing beyond the above; the remaining changes in this release are diagnostic logging improvements (for troubleshooting dropped/lost recordings on wireless keyboards) that don't change app behavior, plus internal documentation updates.

Fixes

• Windows: fixed the recording prompt appearing off-screen (invisible) when moving between monitors with different DPI scaling — e.g. dragging from a 150%-scaled display to a 100% 1080p display could push the prompt below the visible screen. The prompt now positions itself correctly on the monitor under your cursor regardless of scaling differences between screens. (macOS and Linux behavior is unchanged.)

What's new

• Tap-to-recover, double-tap-to-lock hotkey: A quick, short press of Ctrl+Shift (under 500 ms, with no second press) is now a "tap." If a failed-dictation card is showing, tapping re-inserts that card's text at your cursor instead of discarding it — no need to reopen History and copy/paste. Without a card showing, a tap behaves as before (a mis-trigger cancel). A second quick press within 350 ms instead locks the recording hands-free: it keeps recording without holding the keys down, until you press again, hit Escape, or 12 minutes pass. A tap followed by holding the keys is treated as a normal dictation, and long-hold recordings show a short tip about the double-tap feature (shown at most three times).
• Note for Windows users: switching keyboard layouts with Ctrl+Shift twice in quick succession can accidentally trigger a lock.
• Filler-word removal: Final transcriptions can now automatically drop hesitation fillers (嗯, 呃, and standalone 额) along with the surrounding pause punctuation, e.g. "觉得，呃，我们" becomes "觉得，我们". Fillers that are quoted, listed alongside other interjections, named, used as parts of words, or reported as someone's spoken answer are left untouched. A new "Remove filler words" setting controls this and is on by default; results that were only filler become empty and are treated as no speech.

Fixes

• Filler removal no longer strips 嗯/呃 when they appear inside quotation marks (“”, ‘’, 「」, 『』, or straight double quotes), since quoted speech should stay verbatim (e.g. 他说："嗯，好吧。" keeps its 嗯). Unmatched quote marks (like an apostrophe, or a dropped closing quote) are not treated as quoting.

Fixes

• Fixed emoji and other special characters occasionally arriving as garbled text when typed into the focused app — long insertions are now split only between whole characters, so multi-part characters (like emoji) never get cut in half.
• Fixed inserted text sometimes triggering unintended shortcuts when a modifier key (Shift, Control, Option, Command, Caps Lock, Fn) was held down while dictation was inserting text.

Improvements

• Faster text insertion: the pause between chunks of typed text was reduced from 5ms to 3ms, making long dictations appear noticeably quicker.

Improvements

• Redesigned the engine picker on the Home screen: instead of three tall buttons plus a separate row for Qwen's size, it's now a single compact row of pills. The Qwen pill shows its size directly (with a quick switch if both 0.6B and 1.7B are downloaded), and a "More" menu lets you jump straight to Groq or Nemotron.
• Local engines are now marked with a computer icon and cloud engines with a cloud icon, so it's clear at a glance whether your speech leaves your machine.
• Engine buttons now only show a status label when something needs your attention (e.g. not downloaded yet), instead of always showing a status.

What's new

• Settings can now suggest switching Qwen to the 1.7B model based on your own machine's measured dictation speed, instead of guessing from the chip name. After enough dictations on 0.6B, if 1.7B would still respond quickly, SayType suggests it in the Qwen details; if 1.7B turns out to be slow on your machine, it suggests switching back to 0.6B.
• Home and Settings now treat Qwen as a single engine with two sizes (0.6B / 1.7B). Switching away to another provider and back to Qwen restores whichever size you used last, and once both sizes are downloaded, Home shows a quick 0.6B/1.7B toggle.
• "Download 1.7B" keeps dictation working on 0.6B while it downloads, then switches to 1.7B automatically once it's ready — unless you changed engines while it was downloading, in which case it stays put.

Improvements

• Onboarding no longer offers the 1.7B model up front; first run now only asks local vs. cloud, keeping setup simpler.
• Groq and Nemotron are now tucked under "More" in onboarding, Home, and Settings, while whichever engine you're currently using always stays visible.
• Qwen 1.7B is now described as a "larger" model option rather than "experimental."

What's new

• The Dictionary page now uses chip-style tags instead of a textarea: type a word and press Enter (or comma, or paste a comma-separated list) to add it, and click the × on any chip to remove it. Changes save automatically — no Save button needed.

Improvements

• Rewrote a lot of the Chinese (and some English) UI copy to read naturally instead of like translated implementation detail — clearer wording around API keys, transcription, and engine names.
• The local model delete confirmation now shows the model's actual download size instead of an inaccurate fixed "~1 GB".
• The Nemotron latency setting is now labeled "faster text" / "more accurate" instead of showing internal timing numbers that didn't match what users actually experience.
• The Accessibility permission explanation no longer claims SayType "never uploads anything" (untrue when using cloud engines) — it now just explains what the permission is used for.
• Various error/status messages were corrected to better match actual behavior (e.g., re-transcribe failures, GPU fallback persisting until restart, integrated-GPU slowdown notes).

Fixes

• None beyond the copy corrections above.

What's new

• Nothing new was added this release; the main change is a removal (see below).

Improvements

• Removed the translate mode (hold Shift+Alt to get English text) — it was rarely used and added its own provider picker and upload-consent prompt. Shift+Alt is now just a regular recording shortcut. If you relied on this, please open an issue.
• A local engine (Qwen/Nemotron) now simply hides the API key field when it's not needed, instead of parking it inside a separate translation panel.

Fixes

• An update-check failure used to dump a long raw error message into the Settings status pill, squeezing the "Software updates" text into a single word per line. It now shows a short "Couldn't check for updates. Try again later." message, with the full detail available in the tooltip and the log.
• Fixed a release-pipeline timing issue where Mac users could briefly see an update error in the minutes right after a new version was tagged, because the release could go public before the macOS build had finished uploading. Releases are now only published once every platform's build is ready.

Fixes

• Fixed the model picker in Settings sometimes showing or saving the wrong model when you switched engines or picked a model while a previous choice was still being saved. In particular:
• Reopening the currently active engine's drawer while a save was still in progress could show a stale model, so clicking the model you actually wanted looked like a no-op and got ignored.
• A model chosen right as you switched to a new engine could be silently dropped once the switch completed.
• Opening a different engine's drawer while a pick was still saving could cause that pick to be lost.

These could result in the wrong transcription model ending up selected and saved without any indication something went wrong. Model selection now correctly reflects (and honors) the most recent pick in all of these cases.

Fixes

• Fixed a bug in Settings where quickly switching cloud models (e.g., picking Large V3 then Turbo again before the save finished) could leave the wrong model highlighted, or cause an older pick to be re-saved later. The last model you pick now always wins, even if a save was already in progress; if a save fails, the highlight correctly falls back to the model still in use.

Improvements

• History now keeps the last 200 dictations instead of 100, so recent activity spans about a weekend of use instead of about 42 hours.
• Rewrote the Qwen3-ASR 1.7B model description in Settings so it tells you what matters as a user (download size, memory use, and speed) instead of reading like a benchmark log.

What's new

• Engine selection in Settings now works like a Wi-Fi list: clicking a ready engine (key entered, or model downloaded) switches to it and opens its details in one click. Clicking an engine that still needs a key or a download just opens it, with a hint like "Not in use yet. Add an API key, then click OpenAI again to switch."
• The expand arrow on each engine row is now a separate button, so you can look at (or edit) another engine's settings without switching to it.
• Picking a model inside the currently active cloud engine's drawer now applies immediately — no more separate "Use <model>" button.
• If activating an engine fails, the error now stays attached to that engine instead of jumping to whichever drawer you open next.

Improvements

• The engine that's actually in use is no longer dimmed, even if it's an experimental one.
• The Qwen3-ASR 1.7B row is now a single line; its cost details moved into its drawer.
• When Qwen is active, the language setting shows a simple one-line summary instead of a disabled dropdown with a boxed note.
• The "in use" indicator is now a clearer circle-with-checkmark, and the outdated "click a title for details" instruction was removed from the page.
• Notices like "Using this cloud model uploads dictation audio to its provider" are now shown as plain text instead of looking like an input box.
• Tightened spacing throughout the Settings page so drawers and cards take up less vertical room.

Fixes

• Fixed the expand chevron on engine rows being invisible — it was rendered outside the visible row area and never showed up.
• Fixed drawer contents (API key, model choices, notices) not lining up with the engine's name above them.
• Fixed a bug where switching between cloud engine drawers could scramble the order of the API key field, model picker, and notice.

Fixes

• Fixed a visual bug in the dark theme where the transcription engine list in Settings looked washed out with a milky grey fog, especially around the expanded engine details and selected/experimental rows. Row dividers and backgrounds now render correctly on both dark and light themes.

What's new

• Redesigned dark theme. The old dark theme was a mid-tone blue-purple gradient that didn't actually read as "dark" — panels barely stood out and white text had weak contrast. It's now a proper near-black surface built from the same warm ink and eucalyptus-green accent as the light theme, with depth conveyed through lightness (recessed sidebar/inputs, raised cards). Native dropdowns and scrollbars now follow the dark theme too, and the recording prompt glows in the app's accent color instead of indigo.
• "What's new" link on the update card. Once a new version is downloaded, the Home update card now includes a link that opens that version's release notes on GitHub, instead of just saying it's ready to restart.
• Easier fix when Accessibility permission won't show up. If SayType never appears in System Settings' Accessibility list, Settings now offers the same "Reveal SayType in Finder" button and drag-in hint that Home already had, so you can drag the app in manually. Opening the Accessibility pane from Settings also shows the drag helper immediately.

Fixes

• Corrected the "Start minimized" description: it applies to every launch (not just system startup), and a manual launch from the Dock also stays hidden for the first few seconds — the setting text was misleading about when it applies.

Fixes

• Windows: microphone status is now detected correctly. SayType tracks whether the last recording attempt actually succeeded and reports a clear status — ready, denied, no device found, or error — instead of assuming the microphone is always available. Denying access, opening the Windows microphone privacy settings, and retrying now works as expected, and Home/onboarding/Settings all reflect the current state.
• Windows: "Copy to clipboard" now works reliably. Manual copy uses a native Unicode clipboard write, and if the copy fails you'll see a visible error instead of silent failure.
• macOS microphone/accessibility permission behavior is unchanged.

What's new

• Added a "Merge spelled-out letters" option in Settings (on by default). When enabled, independent uppercase English letters separated by ordinary spaces — like "A P I" — are automatically joined into "API" in the final transcript. This applies once a dictation is fully assembled, so History and insertion use the same formatted result; live previews and in-progress chunks are left untouched. Punctuation, tabs, and line breaks still act as normal separators and won't be merged.

What's new

• Redesigned first-run setup: a clear step-by-step guide (Welcome → Privacy → Engine → Microphone → Accessibility on macOS → Practice) that you can leave and resume from the same step later via Settings.
• New installs now default to the local Qwen3-ASR 0.6B engine on all platforms; machines detected as low-end (under 8 GiB RAM or fewer than 4 CPU cores) default to OpenAI instead. Existing installs keep whatever engine you already had configured.
• You can now pick a local model as your engine even before its download finishes — the app tracks readiness separately from your selection, and the download continues in the background while you finish permissions. It won't start downloading anything, or switch your active engine, on its own.
• On supported Apple Silicon (M4/M5 and newer with 16 GiB+ RAM), onboarding shows a side-by-side comparison (speed, memory, download size) for the larger 1.7B Qwen model as an option; 0.6B remains the recommended default everywhere.

Improvements

• Dictation now sticks to the transcription engine (local/Groq/OpenAI) that was active when you started recording — for the whole session, including automatic retries — even if you change the engine in Settings mid-recording. This avoids silently sending later chunks of the same recording to a different provider than the first ones.
• The practice step in onboarding now shows a more compact status readout.

Fixes

• Changing the dictation language in Settings now applies to Nemotron live dictation immediately. Previously it only took effect after restarting the app.

Fixes

• Windows/Linux: microphone is no longer held open between dictations. Previously the app kept a single microphone stream open for the whole session, which left the OS "microphone in use" indicator on constantly and could keep Bluetooth headsets stuck in hands-free (call-quality) mode. Each dictation now opens and closes its own microphone stream. (macOS is unaffected — it already used a different, native capture path.)
• Onboarding: fixed the final practice step not accepting keystrokes. The last onboarding page asks you to try dictating a sentence, but the text field wasn't focused, so typed/dictated text could be silently ignored. The field now receives focus automatically when that step becomes ready.
• Settings: language selection now works for the Nemotron engine. It was previously locked to automatic language detection like Qwen; you can now choose a specific language for Nemotron, and your saved language choice is preserved when switching or inspecting engines.

Improvements

• Simplified the engine-switching feedback in Settings: removed the Undo button and redundant success message, and hidden the activation button for the model already in use. Selection indicators, cloud-upload notices, and failure messages remain visible where relevant.

What's new

• Add optional Qwen3-ASR 1.7B Q8 local transcription. Qwen 0.6B remains the recommended default; 1.7B is experimental, with clear download and memory-cost information.
• Keep both Qwen model sizes independently downloadable and removable. Switching sizes retires the previous inference worker.
• Separate engine selection from configuration in Settings: use the check button to activate an engine, and its title to expand or collapse details. Browsing models, editing API keys, and downloading assets do not silently change the active engine.
• Show the active model, support undo and preserve the previous selection if activation fails.
• Improve Home engine availability/progress indicators and keep Settings saves coordinated with Home quick switching.
• Make engine checkmarks smaller while preserving their click area.

Windows GPU acceleration continues to use the optional Vulkan package. No CUDA package is added. The larger model's performance depends on hardware and available acceleration; Windows/Linux real-device performance is not newly certified by this release.

Improvements

• Added end-to-end sample coverage checks for chunked local dictation. Before inserting the final text, SayType verifies that captured audio was fully accepted, queued, submitted, and completed. A detected coverage mismatch now enters the existing audio/text recovery flow instead of automatically inserting a partial result.
• Added privacy-preserving tail diagnostics: per-chunk sample ranges, request byte counts, result character counts, and separate release/drain timing. These new diagnostics do not log audio or transcript content.
• Legitimate empty transcription results remain valid and do not trigger retries or errors.

This release adds safeguards and diagnostic evidence for investigating intermittent tail loss; the original intermittent trigger has not yet been reproduced or conclusively identified.

What's new

Redesigned Settings

• Two focused tabs: Dictation Settings and App Settings.
• Engine-specific accordion drawers keep API keys, model choices, and local-engine controls beside the engine they belong to.
• Settings save automatically, with visible save and permission-check feedback.
• Segmented language/model controls, visual theme previews, and rounded startup switches.
• Cloud model cards show prices and tradeoffs; a single available model is presented as information rather than a selector.
• Qwen and Nemotron report their own download/readiness state independently.
• Optional cloud translation is separated from local dictation, with upload consent and an explicit cloud provider.
• Clearer spacing, an improved version/update control, and OpenAI listed before Groq.

Reliability fixes

• Cancelling a translation also clears its pending consent prompt.
• Missing local models no longer block unrelated settings changes.
• An explicitly selected translation provider never silently falls back to another provider when its API key is missing.

Fixes

• Failed recordings remain available for retry, including when a retry returns no text. History shows the latest failure reason.
• Automatic retries reuse pending History entries. If a manual retry finishes first, a later automatic result is saved separately so the text returned for insertion remains available in History.
• Fixed missing History entries for failed cloud/translation requests with incomplete capture. Recovery ownership now follows the recording-start provider snapshot, preventing gaps when Settings change during recording.
• Added localized retry errors without internal file paths, and made History search match the displayed error text.

Fixes

• Failed transcriptions no longer disappear without a trace. When transcription fails (missing API key, network error, timeout, etc.), the recording itself is now kept alongside the error message, so you can retry it later instead of just seeing a lost error entry in History.
• Retrying a failed recording ("re-transcribe") no longer re-runs it on whatever provider failed before — it uses whatever engine you have configured now, so fixing your API key, network, or switching providers actually fixes the retry.
• A single recording that gets auto-retried (e.g. after a timed-out upload) now produces one History entry instead of two, and a retry that later succeeds correctly replaces the failed entry rather than leaving an orphaned failure row next to the successful text.
• Retrying a failed entry now shows an up-to-date reason if it fails again, instead of leaving a stale error message that no longer matches what actually happened.
• Clearer error messages for edge cases: a recording whose audio has since been deleted, a recording whose format the local engine can't read (with guidance to switch to a cloud engine in Settings), and other unreadable-recording cases.

Improvements

• Failed recordings are automatically cleaned up along with the rest of History (subject to the same 100-entry cap), so they don't accumulate indefinitely on disk.

Recording reliability

• Bound native microphone startup and shutdown waits so a stalled stop cannot indefinitely block subsequent dictations. The frontend allows 8 seconds around the backend's 5-second shutdown deadline.
• Prevent a failed startup from displaying Listening. Restore WebKit fallback when native capture fails and device release is confirmed; report busy devices and startup timeouts more precisely.
• Continue collecting Qwen chunked and Nemotron streaming results when recovery-WAV finalization fails.
• Mark interrupted or incomplete recordings explicitly, preserve available text in History, and offer Copy instead of silently inserting a truncated result.
• Add stable native error codes, final capture-integrity diagnostics, and cross-language timeout/error contract tests.

Validation and known limits

• Automated checks: 196 Node tests passed; 164 Rust tests passed, 3 ignored.
• Physical microphone-disconnection scenarios have not yet been verified on real hardware. Windows and Linux installers are CI-built, not real-device validated.
• Interrupted cloud/translation audio remains memory-only; successfully recovered text is saved to History.

Improvements

• OpenAI transcription now uses a single, simplified model: GPT Transcribe ($0.0045/min), replacing the previous three options (GPT-4o Transcribe, GPT-4o Mini Transcribe, Whisper-1) in the model picker. It's cheaper than the two higher-priced models and produces cleaner Chinese punctuation (proper full-width ，。 and correct sentence segmentation) than GPT-4o Mini Transcribe.
• The default OpenAI model is now GPT Transcribe everywhere the app picks a default (new installs, onboarding, and the empty-model fallback).
• Whisper-1 is still used automatically for translate mode, since OpenAI's translation endpoint requires it — no action needed there.
• Old history entries created with a retired model (GPT-4o Transcribe, GPT-4o Mini Transcribe, Whisper-1) still display their original model name correctly.
• Updated in-app text to reflect that automatic Chinese punctuation seeding applies to Whisper models and is skipped for all GPT-family models (previously only mentioned GPT-4o).

Fixes

• Fixed quiet/garbled audio at the start of dictation on macOS. The first ~3 seconds of every recording were captured about 32 dB too quiet, which could clip or lose the start of what you said. Audio capture on macOS now goes through the system's native CoreAudio path instead of the browser's microphone API, which doesn't have this problem — including on the very first dictation after launching the app.
• Fixed a rare case where starting a new recording immediately after releasing the hotkey from a previous one could fail to start cleanly.
• If native audio capture ever fails to start, SayType now falls back automatically to the previous capture method instead of losing the dictation.

Improvements

• The microphone indicator in the macOS menu bar no longer stays lit between dictations — it now only lights up while you're actually recording.
• Bluetooth headsets are no longer forced into low-quality call-audio (HFP) mode while dictating on macOS.

*(Windows and Linux capture is unchanged in this release.)*

What's new

• Local transcription (Qwen3-ASR / Nemotron 3.5) is now recommended by default on all supported desktop platforms, not just Apple Silicon — the onboarding wizard, "recommended" tags, and engine switcher now steer every user toward the no-key local engine first, with Groq/OpenAI as optional cloud fallbacks.

Improvements

• Long dictations using the local Qwen engine are noticeably faster: previously each chunk of a long recording paid a fresh model load, but now one model instance stays warm for the entire hotkey hold, so only the first chunk pays the load cost.

Fixes

• Fixed a rare case where the local Qwen engine's speed optimization could leave a stray transcription process running in memory longer than intended when a recording was cancelled mid-decode; it's now reliably cleaned up.

Fixes

• Fixed the input prompt window continuing to animate after it was hidden, which kept the renderer and GPU busy in the background even when no window was visible.

Improvements

• The input prompt now respects your system's "reduce motion" setting — the breathing glow and insertion-failure pulse are disabled (the amber warning ring still shows, just without movement).

What's new

• Nemotron local transcription now runs on Windows (x86_64, CPU), not just Apple Silicon Macs.

Improvements

• Nemotron's runtime is now downloaded directly from NVIDIA's official release and verified by checksum, instead of being bundled inside the app — this shrinks the installer and lets the runtime be updated independently of SayType.
• Upgrading from an earlier version automatically cleans up the old bundled Nemotron runtime, freeing roughly 10 MB of disk space per machine.

Fixes

• Fixed a bug in local (offline) transcription where a worker process could occasionally hand back the *previous* dictation's text instead of the current one; on Windows this could crash the process outright. The app now detects this and re-runs the transcription cleanly.
• Fixed local transcription failing to start for some users — a leftover OpenSSL dependency from the previous custom-built runtime shipped broken on Windows and could silently fail on macOS if Homebrew's openssl@3 wasn't installed. Local transcription now runs on upstream's official llama.cpp builds instead.
• Fixed long dictations (recorded in chunks) reloading the local speech model between each chunk, which had reintroduced the slowdown chunked transcription was designed to avoid.
• Fixed a case where releasing the dictation hotkey right as a chunk finished decoding could leave an unused ~1.1 GB local-model process running in the background.
• Diagnostic logs for local transcription no longer contain any of your dictated text — only an anonymous fingerprint used to spot repeats, never the words themselves.

Improvements

• Local transcription now starts loading its model as soon as you press the hotkey and releases it right after each dictation finishes, so no memory is held in the background between dictations while the model-load time still stays hidden behind your speech.

Fixes

• Fixed local (offline) speech recognition failing to start on Windows with a "libssl-3-x64.dll was not found" error. The bundled runtime no longer depends on OpenSSL libraries that may not be present on a user's machine.
• Fixed the same underlying issue on macOS, where the local ASR runtime silently depended on a Homebrew-installed OpenSSL library (/opt/homebrew/opt/openssl@3) that most users don't have — local transcription could fail to start without it.
• Transcription quality and speed are unchanged; this release only removes a hidden dependency that could break local ASR on a clean install.

Fixes

• macOS: Switching to SayType with Cmd+Tab now restores the main window when all SayType windows are hidden.
• Existing visible windows, including the recording prompt, are left undisturbed. Activations during the first three seconds after startup are ignored to preserve start-minimized behavior.

More reliable dictation completion

• Local Qwen chunked dictation and Nemotron streaming now finalize on key release without waiting for the recorder's stop callback. Late or duplicate callbacks cannot insert the same result twice.
• Whole-clip dictation gets a longer recorder-finalization grace period, while stalled transcription stages have explicit deadlines and cancellation cleanup.

Better failure recovery

• Failed-insertion and partial text are preserved for recovery, with acknowledged, idempotent History saves. Recovery cards auto-hide and no longer block later successful dictation or repeatedly display old text.
• Late local recording data can be saved for manual retry even when the recorder's stop callback never arrives. Unacknowledged text remains in memory after its card is dismissed; quitting before successful persistence can still lose it.
• Lifecycle diagnostics distinguish actual recorder callbacks from logical completion, without logging transcription text. Log storage is bounded to approximately 6 MB.

Validation includes automated regression tests, real Qwen/Metal smoke tests, and isolated WKWebView tests with synthetic audio and injected missing/late events. The original intermittent incident has not been naturally reproduced. Windows/Linux installers are built in CI; desktop recording/insertion on those systems has not been manually verified for this release.

What's new

• Local Qwen dictation now splits long recordings into chunks in real time while you're still speaking, so only the last little bit of audio needs to be processed after you release the hotkey. Multi-minute dictations that used to take tens of seconds after release now finish almost instantly, with the floating window showing already-finalized text while the rest is still being captured.
• Qwen now starts warming up as soon as you begin recording, cutting the wait after you release the hotkey even further.
• Settings now recommend the Qwen model and show expected latency for Nemotron, making it easier to pick the right local model.

Improvements

• If live chunking can't be set up (e.g. unsupported audio capture), dictation automatically falls back to the previous whole-clip transcription instead of failing.
• If one chunk fails to transcribe, the rest of your dictation is still delivered — you just get a small gap instead of losing the whole recording.
• A failed history save can no longer cost you the transcribed text; the text you see typed and the saved history are now guaranteed to match.

What's new

• Added a second local transcription engine: Nemotron 3.5 ASR Streaming, which runs entirely on-device on Apple Silicon and produces live, real-time transcription as you speak.
• The local engine is no longer a single "Local" option — you can now pick Local · Nemotron 3.5 or Local · Qwen3-ASR independently, each with its own one-time download, in onboarding, the Home engine switcher, Settings, and the tray menu.
• Nemotron is offered as the recommended default local engine; Qwen3-ASR remains available for batch-style transcription.

Improvements

• Tuned Nemotron's streaming recognition profile for its trained 560 ms accuracy window, improving transcription accuracy.
• Tray and Settings menus now show which specific local model is active (not just "Local"), and each engine's download prompt opens to the right model automatically.
• Settings/tray now switch engines without ever landing on an unusable, undownloaded backend — attempting to select an engine whose model isn't downloaded opens its download panel instead.

Fixes

• Fixed microphone-permission detection incorrectly reporting access as already granted before you ever approved it. Requesting the microphone now properly triggers the system permission prompt, with a link to open system settings if access needs to be enabled manually.
• Cancelling a recording/transcription no longer risks cancelling an unrelated one that happens to be running at the same time — cancellation is now scoped to the specific recording session.
• Hardened how settings are saved: concurrent updates (e.g. saving preferences, finishing onboarding, switching providers, editing the dictionary) can no longer silently overwrite each other's changes.
• Fixed an edge case where toggling "launch at login" could leave the saved settings out of sync if applying the login-item change failed.
• Improved recording-lifecycle handling to reduce edge-case glitches around starting/stopping dictation.
• Local (on-device) transcription now automatically cleans up leftover legacy model files.

Fixes

• Fixed local (offline) transcription failing on macOS with an unhelpful "resident llama-mtmd-cli did not reach its initial prompt" error. The bundled speech-recognition runtime was built with a path that only worked on the build machine, so it silently failed to load on user machines. It's now packaged correctly and verified to run standalone.
• If local transcription's fast path still fails to start for some reason, SayType now falls back to its slower one-shot path instead of failing the dictation outright, and the real error is now recorded in History for diagnosis.
• The app now detects when an already-installed local runtime is out of date and re-installs the corrected version automatically, so users who already hit this bug get fixed on update rather than staying stuck with the broken files.

Fixes

• Fixed a bug in the local (on-device) speech-to-text mode on Apple Silicon Macs where, after the background transcription process stayed running between recordings, it could occasionally reuse leftover audio state from a previous dictation instead of fully resetting — which could cause incorrect or repeated text in a later transcription. The local ASR runtime for Apple Silicon has been updated to fully clear its audio/session state between recordings, so each dictation is transcribed independently and correctly.

What's new

• Settings is no longer a separate window — it's now a page inside the main window, organized into three tabs: Voice Input, Transcription, and App.
• Settings → App has a new collapsible "Diagnostic logs" panel that shows the app's log file (size, last-modified time, and a "Copy all" button) to help troubleshoot issues.

Improvements

• Recording can start faster when the focused app is slow to respond: the accessibility-position lookup now times out after 50ms instead of 250ms and falls back to positioning near the cursor, so a sluggish app adds much less delay before the recording prompt appears.
• Startup timing (native and frontend phases of hold-to-record) is now logged internally, with a warning logged if starting recording takes longer than 500ms — this makes future "slow first word" reports easier to diagnose.
• Settings now warns before discarding unsaved changes when navigating away, and shows a clearer "Saved" confirmation after saving.

Improvements

• Local speech-to-text is noticeably faster: the on-device model now stays warm between dictations instead of reloading each time, cutting the delay before transcription starts.
• The local model automatically refreshes right when you start dictating, so the first result after an idle period is more reliable.
• Local transcription requests are now processed one at a time, avoiding overlapping requests that could previously cause errors or inconsistent results.
• Small performance tuning to reduce unnecessary overhead during local inference.
• Minor visual polish to the floating recording prompt overlay.

Fixes

• Fixed the live transcription preview bubble appearing visually behind the recording box, with its shadow bleeding over the text. The bubble now correctly reads as the front-most element while it's shown.
• The preview window is now taller, so the streamed transcription text can show up to three lines (previously two) before scrolling.

Fixes

• Windows: local (offline) transcription no longer flashes a console window and briefly steals keyboard focus away from the app you're dictating into.

What's new

• On-device transcription is now much more resilient. If a local (on-device) transcription hangs mid-decode, SayType detects the stall, automatically retries it once, and preserves recording order so later clips don't jump ahead of a stuck one.
• If a hung transcription still can't be recovered after the retry, the recording is no longer lost — it's saved to your history as a "pending" entry you can tap to retry at any time.

Improvements

• Faster on-device transcription on Windows: an unnecessary warm-up pass before each transcription is now skipped, shaving noticeable time off every clip.

Fixes

• Fixed a Windows-specific bug where voice input could fail to reach the app correctly due to an overly strict content-security policy that didn't recognize Windows' IPC transport, causing audio to be mis-encoded.

What's new

• Added a new drag-and-drop helper that appears when SayType asks for Accessibility permission. Instead of hunting for SayType in Finder, you can now drag its icon straight from a small floating "cloud" into the System Settings permission list.

Improvements

• Redesigned the drag helper's look: a soft, cloud-shaped window with a grounding shadow, replacing the earlier flat white card.
• Closing the drag helper before granting permission no longer leaves the main window stuck on "Waiting for permission…" — the guide button now correctly reappears so you can try again.

Fixes

• Fixed a hard black outline/ring that showed around the drag helper's cloud shape.

Fixes

• Fixed local (on-device) transcription crashing immediately on a fresh install. The app's downloader was dropping required library symlinks when unpacking the local speech-recognition engine, which caused it to abort on the very first transcription attempt. New downloads now extract correctly.

Improvements

• The dictation prompt now appears on the screen you're actually working on, instead of always showing up on the primary monitor. On multi-monitor setups, it follows the app you're typing into (or the mouse pointer if that can't be determined), so it no longer feels like it pops up in a random spot.
• Fixed the prompt occasionally flashing on the wrong screen for a split second when switching monitors.
• The prompt's spacing from the bottom of the screen is now consistent across displays with different resolutions/scaling (e.g., Retina vs. standard external monitors), where before it could look closer to the edge on high-DPI screens.

What's new

• Interrupted local-model downloads can now be discarded, not just resumed — no more being stuck with up to ~1 GB of leftover partial files with no way out but finishing the download.

Improvements

• Notifications (toasts) now stack neatly in a column instead of overlapping when multiple appear close together.
• The local-model download progress bar now matches the app's theme instead of showing the browser engine's default blue.

Fixes

• Fixed a bug where a slow or failed transcription from a previous recording could interrupt and tear down an actively in-progress new recording — the app could stop listening or lose queued text mid-dictation. Transcriptions now only affect the UI/session they actually belong to.

Fixes

• Fixed a rare race condition where two transcriptions finishing at the same time could silently drop an entry from your history.
• Fixed "Launch at login" on macOS: if enabling/disabling it failed, the setting could get stuck out of sync and a later retry would be silently skipped.
• Hardened API key handling so your Groq/OpenAI keys are only ever sent to the Settings window, never to other app windows.

Fixes

• Fixed the live transcription preview bubble clipping the top line of text mid-letter when the text scrolled past two lines. The hidden top line now fades out smoothly instead of being cut off, making it clearer that there's more text above rather than looking like a UI glitch.

What's new

• Local (on-device) transcription now shows live progress: as the model decodes your recording, the transcript streams into the prompt window word-by-word instead of leaving you staring at a blank "Processing…" for up to ~30 seconds on longer clips.

Fixes

• The transcription preview bubble was being invisibly clipped by its container and never actually appeared — for local *and* the final preview text alike. It's now positioned correctly and always visible, showing the most recent lines.

What's new

• Local, on-device transcription (Apple Silicon Macs): SayType can now transcribe entirely on your Mac using a local Qwen3 model — audio never leaves your machine. On Apple Silicon it's recommended as the default engine; cloud (Groq/OpenAI) is still one click away.
• Quick engine switching: Switch between Local, Groq, and OpenAI right from the menu bar tray (new "Engine" submenu) or from a new Engine card on the home screen — no need to open Settings.
• Updated onboarding: On Apple Silicon, setup now leads with the local model (in-place download with progress) and folds the cloud options behind a toggle. The privacy explanation now leads with "audio never leaves this Mac."

Improvements

• The home-screen engine switcher now lives in its own clearly labeled card instead of being buried inside the status card, with a short caption explaining what each engine means for your privacy (local vs. cloud with your key).
• Picking "Local" before its files are downloaded now opens Settings directly to the download panel instead of silently failing.

Fixes

• Settings no longer shows a stale (red) Accessibility permission status when first opened — it now refreshes automatically when the window comes to front or the permission changes, instead of only updating after clicking "check" manually.
• The dictation hint bubble no longer briefly flashes an incorrect, hardcoded shortcut (and the English word "Alt" instead of macOS's "Option") before your actual configured shortcut loads.

What's new

• On-device (local) transcription — SayType can now transcribe speech entirely on your Mac using the Qwen3-ASR model via llama.cpp, with no API key, account, or internet connection required. Manage the local model from Settings (download, progress, delete), see a model badge showing which engine is active, and switch between local and cloud transcription; translation still falls back to the cloud API when needed.
• Automatic updates — SayType now checks for new versions in the background (once a day) and you can trigger a manual check from Settings. When an update is downloaded, restart from the tray menu to install it.

Nothing in this release is user-visible: it's all internal work — the design spec, implementation plan, and CI tooling for the pipeline that now auto-generates these bilingual release notes. No app behavior, features, or fixes changed for users in v1.3.5.

• Add gpt-4o-transcribe and gpt-4o-mini-transcribe models to OpenAI provider
• Implement dynamic audio format selection (MP4/M4A preferred over WebM)
• Pass actual recording MIME type from frontend to backend for proper file extension
• Fix audio format mismatch that caused "Audio file might be corrupted" errors
• Fix misleading "transcription failed" error when transcription succeeds but text insertion fails
• Remove unreliable "Text insertion completed successfully" console log
• Simplify status logic to avoid false claims about text insertion success

Release Highlights

🎉 New OpenAI Integration - Added OpenAI Whisper-1 as an alternative transcription provider 🎨 Improved Settings UI - Enhanced layout and better organization ⚡ Performance Optimizations - Service caching and code refactoring for faster response times 🏗️ Code Quality - Major refactoring for better maintainability and future extensibility