Changelog

0.7.2 · October 6, 2026

Pasting

• In Microsoft Word and other apps with their own editor, dictation is pasted directly again. Since 0.7.0 it wrongly ended up only on the clipboard with the “No text field” notice

0.7.1 · October 6, 2026

New installer window

• The installer disk image window now shows the VoxaFlow design: dark background with the waves from voxaflow.ai, the logo, a clear instruction in English and German and an arrow from the VoxaFlow icon to the Applications folder
• The mounted disk shows the VoxaFlow icon

New welcome window

• Setup on first launch runs in four steps: welcome, permissions, dictation key, models. With moving waves in the background that swell when you press the dictation key
• The dictation key can be changed right in the welcome window, e.g. for external keyboards without fn
• New: “Open VoxaFlow at login”, switched on by default, so the dictation key works right after every restart. Change it under Settings › General
• With “Reduce motion” in the macOS accessibility settings the waves stand still

0.7.0 · October 6, 2026

New main window

• Instead of the tab bar, a window like System Settings: sidebar with search, colored icons and a header card per section, back and forward. New sections Overview, Inserting, iCloud & Data and Permissions; “Hotkey” is now “Keyboard”
• While the window is open, VoxaFlow shows in the Dock and in ⌘-Tab
• Open at login (Settings › General)

Overview

• Words dictated, time saved compared with typing (typing speed adjustable, 40 words per minute by default), dictations, speaking speed, snippets used, each compared with the previous period
• Corrections split into AI cleanup, dictionary and “by you afterwards”
• Words per day, activity by weekday and hour, top apps. Numbers only, never the dictated text, with its own switch under iCloud & Data

History and copying

• History with search, filter by period and app, groups by day and app icon. Copy by double-click, ⌘ C or the icon in the row, “Copy original”, “Clean up again” in another style, export as text, CSV or JSON
• Retention of 7 days, 30 days, 1 year or forever instead of a fixed 500 entries
• Menu bar: “Paste last dictation again” and recent dictations as a submenu, also after a restart
• If no text field is active (e.g. on the desktop), the text stays on the clipboard and the pill shows “No text field”. Before, it was lost

Shortcuts and quick history

• Your own key combinations for quick history, “Paste last dictation again”, “Process last recording again” and “Add selection to dictionary”, without extra permissions
• Quick history as a floating window like Spotlight: search, ↩ inserts into the app you are writing in, ⌘ C copies, ⌘ ↩ inserts the original
• The last recording stays in memory and can be processed again, e.g. after an error

Dictionary and snippets

• Table with search and sorting, edit by double-click, duplicate terms are merged
• Import and export as JSON or CSV (Excel, Numbers) with a preview “new, updated, unchanged”, merge or replace
• Learning from corrections: when you change a word right after inserting, VoxaFlow suggests it for the dictionary. Only the text field VoxaFlow wrote into is read, never a password field; nothing is added without asking
• Backup and restore of settings, styles, dictionary, snippets and optionally the history in one file

iCloud sync

• Dictionary, snippets, settings and statistics on all Macs, and the history if you want. Stored in the “VoxaFlow” folder in iCloud Drive
• End-to-end encrypted; the key lives in the keychain, further Macs get it through a recovery code. Neither Apple nor VoxaFlow can read the data
• Changes apply per entry, deletions reach all Macs, the same term from two Macs becomes one. If iCloud Drive is off, VoxaFlow says so

Audio

• Optionally mute other audio (music, videos) while recording; afterwards everything is as before
• Fallback microphones in a fixed order when the chosen microphone is not connected

0.6.6 · October 5, 2026

Beta until January 31, 2027

• The menu, the welcome window and Settings › About show “Beta: all features, free until January 31, 2027.”
• After that VoxaFlow says once that the beta has ended and points to voxaflow.ai; the menu shows “Beta ended: get the current version …”. Dictation keeps working

Model download

• An interrupted download of the language models resumes where it stopped on the next attempt, even after a restart; before, it started over
• Every file is checked after downloading, a damaged one is downloaded again. VoxaFlow retries short connection drops by itself
• A downloaded model starts without an internet connection
• Models you already downloaded are taken over, without a new download and without extra disk space
• Clear messages when there is no internet, a server error or a full disk, plus “Try again” in the welcome window, the settings and the menu. When the network is back, VoxaFlow continues by itself

Report a Problem

• New in the menu: “Report a Problem …”. VoxaFlow creates a text file with version, Mac model, settings, log entries of the last 24 hours and the latest crash report, and opens an email to hello@voxaflow.ai with the file attached. Alternatively just save the file
• No dictated text, no audio, no history; dictionary and snippets only as a count. Nothing is sent until you click Send yourself

Faster on Macs with M1

• On the M1, long dictations are cleaned up about twice as fast (1.4 to 2.4 instead of 4 to 8 seconds), short ones about 30 % faster. The large model Qwen3 4B is now as fast there as the small one, with the same accuracy

0.6.5 · October 3, 2026

Updates without interrupting windows

• When the daily check finds an update in the background, VoxaFlow no longer opens a window over your work but shows “Install update …” in the menu bar. Only when the check runs right after launch or you check yourself does the update window appear immediately

0.6.4 · October 3, 2026

Automatic updates

• VoxaFlow checks once a day for a new version and installs it on request. New in the menu bar: “Check for Updates …”; under Settings › About: “Check for updates automatically” and “Check now”
• Updates come from updates.voxaflow.ai (servers in Germany), are signed and notarized by Apple
• From this version on updates arrive automatically; from 0.6.3 and earlier install the DMG once by hand

Models from our own server

• Speech recognition and language models come from models.voxaflow.ai instead of Hugging Face, in fixed versions
• Models you already downloaded stay valid, nothing is downloaded again

0.6.3 · October 3, 2026

Long dictations: processed while you speak

• From 8 seconds of recording on, VoxaFlow recognizes and cleans up finished sentences while you keep talking. After you release the key, the text usually appears after about a second, even after a minute and a half of dictation (before 4 to 5 seconds)
• If you correct yourself at the start of a new sentence (“no”, “sorry”) or continue a list (“secondly”), that sentence is cleaned up together with the previous one
• Fixed: periods at pauses in the middle of a sentence and a dropped subordinate clause

Filler words

• After the AI cleanup VoxaFlow removes remaining filler words (uh, um, erm, hm and others; “um” only in English text). An inserted filler word takes its commas with it
• Fixed: without a language model, “um” was removed in German too (“um 16 Uhr” → “16 Uhr”)
• “nee” counts as a correction word

0.6.2 · October 3, 2026

Fixed microphone

• Fixed: with a fixed USB microphone (e.g. Jabra Evolve Link MS) every recording stopped after about half a second with “No microphone found”

Start sound only when really recording

• Bluetooth headsets deliver only silence at first while switching. The start sound now plays only once the microphone actually records, so the first word is not lost

More accurate models

• Large model: a more accurate version of Qwen3 4B, with the same size and speed
• Small model: Qwen3.5 2B replaces Qwen3 1.7B, more accurate, faster on long dictations and with less memory
• On the first launch after the update VoxaFlow downloads the new model (2.3 or 1.7 GB)

Faster short dictations

• Short dictations are cleaned up about 0.1 seconds faster

Small model for Macs with little memory

• New: a second, smaller local model, about twice as fast but somewhat less accurate
• Without your own choice the model follows the memory: under 12 GB the small one, otherwise the large one. The large one stays selectable; on Macs with little memory the settings and the welcome window point out that recognition can get slower, with a button to switch
• In Low Power Mode the AI cleanup takes about twice as long; the guide now says so

Command mode: key order no longer matters

• Fixed: pressing fn first and then ⌃ Control gave a normal dictation, and the spoken instruction replaced the selected text
• If ⌃ Control follows fn within a second, the recording switches to command mode (the pill turns purple)

0.6.1 · October 1, 2026

Permissions stay stable

• Permissions are kept across updates and reinstalls
• Input Monitoring is no longer needed; VoxaFlow needs only the microphone and Accessibility
• New button “Repair permissions”: removes stale entries and requests the permissions again; working permissions stay untouched
• If VoxaFlow was installed with a different signature, the app explains it instead of just saying “not active”

Notarized by Apple

• VoxaFlow is signed with Developer ID and notarized by Apple and opens without a Gatekeeper warning

0.6.0 · October 1, 2026

Settings reorganized

The General tab had grown too long. The settings now have eight tabs:

Tab	Contents
General	App language, clipboard, permissions
Hotkey (new)	Dictation key, usage hint, key check with conflicts and key test
Audio (new)	Devices (microphone, sound output), sounds (on/off, start sound)
AI (new)	Speech recognition (dictation language), cleanup (model with recommendation), style per app, local models
Dictionary, Snippets, History, About	unchanged

• The former Style tab is now the “Style per app” section in the AI tab
• The menu bar opens the matching tab: “fn key is assigned twice” → Hotkey, “Set up permissions” → General

Guide

• New chapter “Settings”: overview of the eight sections and the menu bar items
• “Troubleshooting”: new sections “The text appears in the wrong language” and “The start sound is missing or annoying”
• “?” help also for app language, audio devices and start sound

0.5.0 · October 1, 2026

Audio devices

• New in the settings: choose the microphone and sound output, “System default” or a fixed device
• If the fixed device is not connected, the system default applies automatically, with a notice in the settings; when it comes back, it is used again
• Fixed: after switching from AirPods to a USB microphone, VoxaFlow recorded nothing
• If AirPods switch to headset mode when recording starts, the recording continues without losing what was said
• If the microphone delivers no data, VoxaFlow restarts the recording once, then shows “Microphone not responding”
• Tip for Bluetooth microphones: choose the Mac’s microphone so the AirPods don’t switch to headset mode

Start sound

• “Glass” removed, new last option “No start sound” (the start is then shown only by the pill)

Language is kept

• Fixed: a German dictation into a code app such as Claude was translated into English. The cleanup now always keeps the language of the dictation (except in command mode, where translating is allowed)

0.4.0 · October 1, 2026

Bilingual (English and German)

• The whole interface is available in English and German: menu bar, settings, welcome window, pill, messages, guide and guided practice
• The language follows macOS; English applies for all other system languages
• New in the settings: “App language” (system, Deutsch, English), effective after a restart with a “Restart now” button
• “Language” for speech recognition is now called “Dictation language”

Sounds

• Fixed: the start sound was missing on the first press and after a longer pause
• New default start sound “VoxaFlow sound”, a short, soft tone matching the end sound. Before, the start sound sounded like the macOS alert
• Start sound selectable (VoxaFlow sound, Bottle, Purr, Morse, Glass, Tink) with a preview button

Silence no longer produces text

• Fixed: pressing and releasing the key without speaking inserted made-up filler words. Without speech nothing is inserted now, the pill briefly shows “Nothing recognized”
• Silence at the start and end of a recording is trimmed
• Formal emails no longer replace “thanks” with “Thank you”

0.3.0 · October 1, 2026

Faster

• The AI cleanup of typical dictations takes 0.2 to 0.6 instead of 1.0 to 1.4 seconds
• VoxaFlow prepares the cleanup for the active app as soon as you press the dictation key

Qwen3 4B recommended

• Qwen3 4B was faster and more accurate than Apple Intelligence in every test and is now marked as recommended in the settings. With another choice there is a notice and a “Switch to Qwen3 4B” button, also in the welcome window
• Guide: new section “Which language model?” with a comparison

Changed

• The style per app changes only punctuation, capitalization and line breaks, not your wording

Known limitations

• Long dictations still take more than a second

0.2.0 · October 1, 2026

New

• Guide in its own window: 10 chapters with before/after examples, tips and practice fields, in English and German
• Guided practice in 5 steps (dictation, corrections, lists, hands-free, command mode); VoxaFlow ticks off each step automatically
• “?” buttons and tooltips in the settings with an explanation, an example and a jump to the matching chapter
• Entry points in the menu bar (“Guide …”, “Guided practice …”), the welcome window (“Start practice”) and Settings › About

Fixed

• Command mode also works in VoxaFlow’s own windows, such as the practice fields

0.1.1 · October 1, 2026

Fixed

• The fn check wrongly reported “Press 🌐 twice” as a conflict although Apple’s dictation was off

0.1.0 · October 1, 2026

First working version of the Mac app.

Dictation

• Push-to-talk with fn (alternatively right ⌥ or right ⌘), hands-free mode by tapping twice, Esc cancels
• Command mode: ⌃ + fn with selected text rewrites the selection by voice instruction
• Text is inserted at the cursor, the previous clipboard content is restored
• Pill with level meter, menu bar app with its own icon

Local AI

• Speech recognition with Parakeet TDT v3 on the Neural Engine
• Text cleanup with Qwen3-4B-Instruct: filler words, self-corrections, punctuation, tone depending on the app
• Alternatively Apple Intelligence (macOS 26 and later) or simple rules only
• Lists (“firstly … secondly …”) become numbered lists

Personalization

• Dictionary with correction of misrecognitions, snippets, style per app category, local history

Setup

• Welcome window with permissions and model status; detects granted permissions without a restart
• Check whether fn/🌐 is assigned elsewhere, with a warning in the settings, the menu bar and once in the pill
• Live key test for the dictation key

Known limitations

• The AI cleanup takes about 1.1 to 1.4 seconds
• Only the 25 languages of Parakeet
• Not notarized yet, no auto-update