Release notes and version history for WalkAITalkie.
v2.12.5
"Reload Models" now appears exactly when the speech model failed to load, in every app language. Before, it depended on the English words "model" or "load" in the error text
Offline, a failed speech model download says the internet is needed, and the app tries again once the Mac is back online
When Parakeet cannot be loaded and the app uses the downloaded Whisper model, the menu says so
Without Accessibility access, the menu says the hotkey is inactive and opens the right Settings pane
Once the speech model is loaded, the session button no longer waits for the hotkey
Pressing "Retry" after a failed speech model load no longer leaves the hotkey without function behind "Ready"
Russian dictation uses Parakeet in the Automatic setup. Parakeet covers 25 languages again, Whisper the other 13
The speech model download is visible. A missing model shows its size and the download progress before it compiles, and an interrupted download resumes with both at the next launch
Parakeet is 461 MB. Settings, the download status and the free-space check said 315 MB
Offline after an update, the app uses the Whisper model already on the Mac when Parakeet cannot be downloaded
Whisper dictations drop a closing line nobody spoke, such as "thanks for watching" in Japanese, when it stands alone at the end
The menu bar shows "Ready" only once the speech model is loaded and the hotkey works
Auto-stop waits for the first word. In toggle mode it waits at least 8 seconds before you start speaking, then your silence setting applies
A dictation without speech says "No speech detected" instead of a Parakeet error
Turkish dictation uses Whisper in the Automatic setup. Automatic sent Turkish to Parakeet, which does not support Turkish. It now goes to Whisper, like every other language Parakeet does not cover
Auto-stop on silence is part of Free. In toggle mode, the switch and the timing slider in Settings › General no longer require Pro
The paywall in the direct download shows the price Stripe charges: 7,99 € in euro regions, £6.99 in the United Kingdom, CHF 7.99 in Switzerland and $7.99 everywhere else. The Mac App Store build is unchanged, Apple shows its own price
Session notes assign a task to the person it was addressed to: "Anna, please send the offer" goes to Anna, provided the meeting already knows the name. A speaker's own promise stays with the speaker. This applies to the standard Session Notes format in German, English, French and Dutch meetings
Session notes on your Mac still check their task list when the model leaves out the closing brace of its answer, so repeated and invented tasks no longer slip through in that case
Stopping a session no longer waits for a second attempt that cannot help. Only a speech recognition engine failure is retried, so silence or a missing speech model shows its message right away
Two German texts use their umlauts: the Clean Up note in History and the notes language picker
A Dictionary export can be imported again (Pro). Every exported file failed to import because the export wrote dates in a format the import did not read. Import now reads exported files as well as the app's own dictionary file
The Mac App Store build no longer mentions a sign-in it does not have. Settings → Speech & Text now simply says everything on the page runs on this Mac
When session notes merge repeated action items, a task is no longer joined with its own negation: "Do not raise the budget" and "raise the budget" stay two tasks, in the languages Clean Up already guards
The Dictionary counts its terms correctly: "1 term" instead of "1 terms", in all app languages
The unfinished recording card in the menu shows its whole headline instead of being cut off
Settings → General shows the speech language by name, with its flag, as Speech & Text already did
Session notes keep entries that start and end with a bracket, such as "(No date) Anna clarifies this by Friday (CW 42)". They were taken for an empty placeholder and dropped, on your Mac as well as with a custom prompt or a different notes language
Placeholder key terms no longer appear in session notes with a custom prompt, a cloud or CLI provider or a different notes language. Real terms stay
A cancelled product load no longer shows a stale error in the Mac App Store paywall
Dictation through Codex or Claude Code (direct download) works right after the app starts, launch at login included. Before, Clean Up and Translate ran on this Mac after every launch although dictation was opted in. This affected 2.11.0 and 2.11.1
E-mail, Markdown and Copy name the time the session started, in History and in the first result window. A recovered recording is dated by when it was recorded, not when it was recovered. Sessions saved before this version keep their saved time
Clean Up keeps more German cleanups. Speech recognition often writes the hesitation "ähm" as "im"; removing it before an adjective or adverb no longer sends the sentence back as spoken, while the real preposition ("im kurzen Gespräch", "im Büro") stays protected. A corrected amount keeps its currency
Clean Up keeps the good sentences when one sentence fails the safety check. Until now one protected change anywhere sent the whole dictation back as raw text. Now only the sentences with a protected change stay as spoken, the rest is cleaned, and History says how many sentences were kept
New menu item Open Last Session opens the most recent session in History with the same areas and actions as the first result window
Long sessions are transcribed while you record. One worker transcribes the segments already closed at pauses, so Stop only waits for the remainder
Speaker separation shows "usually about …" for recordings of 20 minutes or more. Shorter recordings get no estimate rather than an invented one
On-device session notes merge duplicate tasks: two entries count as one when their words overlap enough and the same person or the same numbers stand behind them. The more specific wording is kept and nothing is rephrased
A dictation's detail view shows the provider that wrote the text, also for a dictation converted to session notes in History
The cleanup hint for the retired Gemma 4 E4B model appears a second time, because anyone who saw the first hint still has the model (5.2 GB) on disk. A retired active model now falls back to Qwen 3.5 4B instead of 2B on 8 GB Macs
The e-mail draft in the first result window uses the same export as History. Try again with the model not loaded no longer overwrites a good stored version with the raw text, and a cancelled run writes nothing
Confirmed speaker names are used when re-summarizing in the first window, and names you typed are committed before Try again and Apply
Two missing History detail strings are translated in all 11 languages
Clean Up and Translate are calibrated from measurements. On a held-out test set of 300 dictations in 25 languages, the earlier settings delivered a good cleanup for about a third of dictations. Deterministic decoding, a clearer Clean Up prompt and a new Translate prompt (fillers and self-corrections are no longer translated) lift the on-device results substantially. Good translations went from 17 % to between 70 % and 77 % on a smaller test set
New installs on Macs with 16 GB or more get Qwen 3.5 9B (5.6 GB), which cleaned 65 % of hard test dictations well against 44 % for the 4B. Macs with 8 GB keep Qwen 3.5 4B. Existing installs are not switched and nothing large downloads on its own; Settings recommends the better model. Gemma 4 E4B is retired
The Clean Up safety check understands all 25 speech languages. It used to throw away almost half of correct cleanups outside German and English as changed wording. It now accepts fillers, hesitations, abandoned starts and corrected values in every supported language, while numbers, dates, amounts, names and negations stay strictly protected
Session notes on this Mac find the tasks: they read the meeting part by part and reconcile the task list. In tests they found 90 % to 96 % of the agreed tasks instead of 25 %. Notes made on this Mac say so and ask you to check who does what
Codex and Claude Code can also clean up and translate your dictation (direct download, Pro, opt-in). A separate checkbox under the selected accelerator; each dictation then uses your own subscription. Only the selected provider receives the text. If it fails, this Mac takes over and the result says who wrote it
Session results reopen from History with the same areas and actions as the first window: the section bar, key terms, warnings, Copy, Markdown and e-mail draft, with the format change above the content. Opening a session never re-runs speech or language models. History rows have a new Open Session context menu item
Assign speakers with text evidence. Each speaker gets a card with up to two passages and timestamps plus a name field, in the first result window and in History. Confirmed names appear in the transcript, notes, action items, Copy, Markdown and the e-mail draft, but only where the speaker is clearly known
The Get Pro and Open Models banner buttons open Settings in front of an open result panel, and so do the menu and Open Settings in the result window
One F7 press ends a dictation promoted to a session, also in toggle mode, and the next Right Option press starts a fresh dictation
A fixed session notes language is visible in the result window and the menu, also for sessions started with F7, and on-device notes follow a pinned notes language for tasks and notes, not only the recap
Recovering an unfinished recording cuts at pauses like the live path
German "ähm" transcribed as "im" no longer makes Clean Up fall back to the raw text. The real preposition ("im Büro", "im Mai") stays protected
Create Email turns session notes into a draft that reads like a normal mail: a dated subject, a greeting, when the session started and how long it ran, the summary and notes under plain headings, numbered action items and a sign-off, written in the language of the notes
When session notes are too long for one email, the action items are shortened last. If no mail app opens, the subject and text are put on the clipboard. Copy now copies what the window shows, including the key terms
Sessions recorded across a room are split at real pauses in speech instead of every few seconds, and a long stretch without a break is cut in a pause. In our tests with recorded meetings this caused fewer word errors in English and German
A dictation promoted to a session is no longer ended early by the silence auto-stop, unplugging a headset during it no longer discards the recording, and the next dictation no longer contains text from the earlier recording
A session that fails to start no longer loses the next dictation, and after a session the dictation auto-stop works again
History → Convert to… applies the chosen format including its translation and notes language, preselects the format the recording was made with, shows the right format afterwards and keeps key terms
Settings → Pro: Activate license again and Deactivate This Mac show an error instead of failing silently
Session notes written in a different language than the one spoken are marked Experimental when summarized on your Mac, because the on-device model can miss or invent action items
Session Notes, Clean Up and Raw keep their English names in every app language
The app no longer quits when Codex or Claude Code has to be stopped after a timeout, and Clean Up in German no longer treats “er” or a sentence-opening “Halt” as filler
VoiceOver reads the buttons in the result windows, such as Create Email and Copy, by their own names again
Session notes tell speakers apart with a new on-device model, NVIDIA Nemotron 3 Diarization. Every word is assigned to the person who said it, so short interjections keep their own label
Speaker labels handle up to 8 speakers and no longer stop at 90 minutes: the recording is read in 30-second pieces, so memory use stays flat however long the session runs
If you already had speaker labels set up, the new model (about 97 MB) arrives automatically in the background; the old one is removed once the new one has loaded
If less than 1 GB of memory is free when a session ends, speaker labels are skipped with a notice. The transcript is always complete
Dictation: the Parakeet engine now runs on FluidAudio 0.17.1, which made fewer word errors in our German benchmark
History shows the real reason when a session-notes summary is incomplete, for example missing action items, instead of reporting a timeout
Stopping a session within a split second of a pause boundary no longer makes the transcription fail
About → Open Source Licenses includes the OpenMDW-1.1 license text and the notices for Nemotron 3 Diarization
Session notes tell two or three speakers apart again on long meetings. Overlap is added up per speaker and a chunk that spans several people is split at the turn boundaries
Dictation no longer picks up an invented closing sentence from the silence after your last word
Clean Up keeps your wording: if the model swaps a pronoun, drops a negation or changes a number, the raw text is kept and History shows why
First-person commitments (“I’ll send the budget on Friday”) land in the action items instead of the notes
You can choose the language your session notes are written in, per format; the default stays the language of the recording
Session notes: the number-consistency filter now runs for German and English only. In other languages, correct sentences with numbers written out as words are kept
Each recording mode in the menu now arms its own format instead of keeping the previous mode's
Settings → Speech & Text: the language list uses the same availability logic as General, so both places offer the same languages
Settings has been rebuilt around one consistent vocabulary. A badge now means the same thing on every page, so you can tell at a glance what is merely suggested, what is downloaded, and what is actually running
The Account tab is now called Pro, and the Processing tab is now Speech & Text. Your licence key goes into Settings → Pro
Speech setup has always been automatic — now it says so. The app shows which engine and model it picked for your language, and the expert options only appear once you take over yourself
Onboarding is down to three steps
Mixed-language recordings stay readable. An internal language hint used to mangle English passages inside a German recording; it now applies only to languages written in a non-Latin script
Session notes no longer invent numbers. Figures that never appear in the recording are removed, and the notes tell you when that happened instead of quietly changing them
Settings reopens on the pane you had open last instead of jumping back to General
In dark mode, text on the Pro surfaces is legible again
Direct download: the cloud accelerator card now appears for everyone with Codex or Claude Code installed, not only when the app was started from a terminal. Detection runs under the same environment as the summaries themselves
Detection is bounded to five seconds and no longer freezes Settings if a CLI does not answer
Settings say what is actually wrong (not installed, not signed in, Node.js runtime not in a standard location, and more) and offer a Check again link
The processing options, including the cloud accelerator, no longer disappear when your dictation format is Raw — which was the default, so most people never saw them
The translation target language is now visible and changeable directly in the menu, and choosing your own speech language as the target is called out instead of silently doing nothing
Cloud processing with your own OpenAI or Anthropic API key has been removed. Speech recognition always runs on your Mac
Direct download, Pro: the cloud accelerator can write session notes through the Codex CLI or the Claude Code CLI you are already signed in to. Text only, off until you choose a provider and confirm it, and the app shows which provider actually wrote the summary
The Mac App Store version has no cloud processing at all
v2.6.0
Pro is now a single one-time purchase with no subscription, credits, or account. Existing licenses keep working
Choose a format before recording from the menu. Free includes Raw and Clean; Pro unlocks the other formats
The speech language picker is now directly in the menu
The Processing tab starts with a short summary and folds advanced engine options away. Choose on-device processing or your own API key
Updated model catalog with TranslateGemma for translation, retired outdated Whisper variants, and verified model sizes
Hardened Pro entitlement checks and removed the retired cloud account
Translation overhaul: custom Formats now reliably apply your chosen target language, Quick Copy respects your active Format, and a new target-language picker in the Formats tab puts translation exactly where you set everything else up
AI model storage moved to Application Support with automatic migration — fixes silent re-downloads and model-load failures caused by iCloud offloading model files from your Documents folder
AI processing problems no longer fail silently: a clear banner and menu-bar state now tell you when the on-device model needs attention, so your text never just disappears
Automatic session summaries: Meeting and session recordings now summarize themselves as soon as the on-device model is ready, with clearer progress and a dependable cancel — capture, transcription, and summarization all run on your Mac, nothing is ever uploaded
Summaries keep speaker context, so who-said-what carries through into your structured meeting notes and follow-up email drafts
Refined Settings: your chosen Format is now the active preset, with clearer Quick vs. Session capture labels and a more discoverable layout
Security hardening for direct distribution: the offline Pro cache is now signed, stored via Keychain-backed state, and rejects tampered or rollback-prone cache data
StoreKit role sync now refuses to report Pro status unless it has a verified transaction JWS, reducing false entitlement updates
Meeting and long-recording reliability improvements, including persistence-error toasts for failed history/status saves
Format and Meeting Mode polish from the record-first hardening cycle, with additional regression coverage across the Swift test suite
New Meeting Mode: record a meeting and get a structured summary plus a ready-to-send follow-up email draft — capture, transcription, and summarization all run on your Mac, nothing is ever uploaded
Smarter on-device language models: refreshed catalog with Qwen3.5 and Gemma-4 for faster, higher-quality text cleanup, translation, and summaries
Cleaner account experience: the Account tab stays out of the way until you opt into cloud features, with one-click Restore Purchases and Delete Account
Upgraded Parakeet speech engine: FluidAudio 0.15.2 with nine months of upstream accuracy and speed improvements, fully ANE-native
Script-aware language filtering keeps recognition locked to your selected language, noticeably reducing wrong-script errors for German and other non-English dictation
Same private, on-device experience — no audio ever leaves your Mac