Changelog

Recordings up to 20 seconds are now processed whole, in a single pass of the speech engine - no internal seams that could cause repeated words or odd add-ons. Longer dictations still use streaming processing, so you never wait for the result.

When the system switches the audio route mid-dictation - for example, another app grabs the microphone - Flash notes where the recording was joined and removes the word duplicated at the seam instead of leaving it in.

The filter for phrases invented from silence now acts only at the very end of a dictation, and only at low recognition confidence. Quietly spoken real words in the middle of a sentence are never touched.

In the double-tap activation mode, stopping a recording could silently start another one listening in the background, whose transcript later landed in your text as unintelligible fragments. This was the source of most “strange add-ons” - fixed.

A smart confidence filter recognizes phrases the speech engine could invent from silence at the end of a recording and removes them before they reach your text. When you actually dictate “thank you”, it stays - the filter looks at recognition confidence, not a word list.

We fixed the seam between chunks of longer dictations - repetitions are caught even when the speech engine wrote the same word slightly differently.

On long meetings, transcription could silently lose several minutes when processing fell behind the recording. Failed segments are now retried, and any loss is explicitly flagged.

In Full capabilities mode, Flash analyzes the flow of the conversation and assigns real names picked up from context (“maybe ask Susan?” → Susan), building the meeting’s participant list. After you correct a name manually, one click applies it across the whole transcript.

Only obvious mistranscriptions - names and cut-off words - are corrected, and built-in safeguards make sure no utterance can be shortened or lost.

“All rights reserved” and similar silence artifacts disappear from meeting transcripts.

Everything in Pro - the best cloud engines, AI correction, and Meetings - for 7 days, with no payment details required. One trial per device.

Alongside your microphone, Flash can now record system audio (macOS 14.4+) - your callers’ voices from Zoom, Teams, or Meet. The transcript is automatically labeled “You” and “Participants”, and smart deduplication removes echo picked up by the microphone.

Flash now waits for AirPods and other BT headphones to fully connect, retrying the audio engine start. Switching or unplugging headphones mid-sentence no longer cuts off the recording - Flash switches over on the fly.

A new activation mode: double-tap Fn to start recording, single tap to finish and paste. Perfect for longer dictations - and you can now pick your own activation key in settings.

Rescue re-sync finds your position anywhere in the script when you go off-text or skip a paragraph, and manual scrolling stays smooth even at several thousand words.

Vibe Coding now works with Cursor, Claude Code, Codex, and GitHub Copilot - in code editors and terminals. Flash locally scans your chosen project folders and spells variable, function, and file names flawlessly, even when you dictate in your native language.

A completely redesigned interface - graphite glass-effect background, translucent cards, and a calm indigo accent - with full light-mode support. The new onboarding walks you through downloading the offline model and granting permissions via drag & drop.

Dictation stays unlimited, forever, and the free offline model (466 MB) stays too. History now keeps your last 30 dictations, the dictionary holds 5 custom entries, and the teleprompter joins the Pro feature set. All your existing data remains untouched.

Utterances are now intelligently merged into coherent text blocks - a new line only appears at a natural pause in the conversation. No more excessive splitting into short fragments.

Flash automatically pauses recording when your Mac goes to sleep and resumes after wake - no data loss or transcript desynchronization.

Transcripts from 2-hour+ meetings are now split into chunks and processed separately, preventing AI cleanup from truncating the text. A progress indicator and cancel option are shown during processing.

A new “Rename speakers” button detects unnamed speakers in the transcript and lets you replace generic labels with real names in one step.

Security improvements in AI text processing, update delivery, and local data storage.

After recording or importing audio, AI automatically transforms the raw transcript into a clean, readable version: merging fragments into coherent statements, identifying speakers, and fixing incomplete words.

Smart silence detection eliminates false fragments and repetitions in transcriptions. Quiet moments in recordings are automatically skipped, significantly improving the quality of the final transcript.

A new "Summarize meeting" button generates a concise summary with key points, action items, and a list of participants. The summary is available anytime via the "Read summary" button.

Fixed app freezing when expanding transcriptions from long meetings. Browsing even 2-hour recordings now works smoothly with native scrolling.

Optimized meeting list refresh during recording — the audio level animation no longer slows down the rest of the interface. Faster search through transcript content.

Before starting a recording or importing audio, Flash checks whether the required transcription model is downloaded. If the model is missing, a prompt appears to download it immediately - no risk of losing a recording.

Meeting transcription always runs consistently in the language set in the app. Smart recognition of technical and industry terms mixed into speech in the native language.

Fixed the Flash app icon incorrectly blinking in the menu bar after finishing a recording.

Double-tap Fn to start recording, single tap to stop. A live audio level indicator and duration are shown in the menu bar. Every recording is transcribed with the highest-quality PRO: Studio model.

Import existing audio files (.mp3, .m4a, .wav, .caf) and get a ready-made transcription. Fast batch processing with the PRO: Turbo model and a notification when it’s done.

One click generates a concise summary with key takeaways, action items, and a cleaned-up transcript with speaker identification. A diff view lets you review changes before accepting.

Precise timestamps every few seconds, seamless chunk stitching with no repeated words, and full markdown support in transcriptions.

Search by title and transcript content, date-range filter, source and AI badges, audio export to .m4a, and a Meetings section in the menu-bar context menu.

Bug fixes and improved stability for the teleprompter scrolling system.