What’s new in VoiceGrep
What changed from version to version. Updates arrive on their own: VoiceGrep checks for them and offers to install the new version in one click.
Version 1.13.0
Tasks have been removed
The “Задачи” (Tasks) section is gone: the “Мои дела” (My tasks) and “Жду от других” (Waiting on others) boards, tasks in meeting, person and project cards, on the dashboard and in exports. The first time the new version starts, saved tasks are deleted, and they can’t be restored. If you need them, copy them down or export your meetings to Markdown before updating.
So that agreements aren’t lost, new meeting summaries include a “Договорились / дальше” (Agreed / next steps) block. If you changed the summary prompt, add the block to it yourself.
Speakers are recognized more accurately
Voices are now split by a new model, “Стандартная (pyannote community-1)” (Standard): it listens to the whole recording at once, splits one person into several “Собеседник” (Participant) entries less often, and recognizes familiar people more often. The cost: short phrases of 1–3 seconds are more often attached to the neighbouring voice; you can fix such a line by hand. The model ships inside the app, nothing to download, and saved voices keep working without any recalculation. To go back, choose “Settings → Транскрипция → Определение говорящих → Прежняя (pyannote 3.1)” (Transcription → Speaker detection → Previous). This version requires macOS 15 or later.
Also, a long reply that the app mistakenly gave to the neighbouring speaker is now re-checked by voice and moves to the meeting participant whose voice it is. This applies to new transcriptions and to “Распознать заново” (Recognize again).
Recording without headphones
Without headphones the microphone hears the other people through the speakers, and their phrases ended up in the transcript a second time — under your name (“Я”, Me). Now these repeats are removed during transcription — only where the audio confirms it really is an echo; your own words spoken over someone else stay. When you play back or export the audio of such a meeting, your microphone is heard only during your replies, so your voice no longer sounds hollow. Recordings made with headphones don’t change. The switch is “Settings → Транскрипция → Убирать эхо собеседников из вашей дорожки” (Transcription → Remove the other people’s echo from your track), on by default. For an older meeting with repeats, press “Распознать заново” (Recognize again) in its card.
Live transcription — automatic, in the meeting card
During a call the conversation text now appears right in the meeting card, in the “Идёт запись” (Recording) block. Live transcription is on for everyone and starts by itself with every recording; to turn it off, use “Settings → Запись → Живая расшифровка во время звонка” (Recording → Live transcription during calls). You can stop and resume it without interrupting the recording using the buttons in the draft’s header.
A separate window is optional: the window icon in the draft’s header or “Живая расшифровка в окне” (Live transcription in a window) in the VoiceGrep menu bar menu. It’s a regular window you can move and resize; the pin keeps it on top of all windows, and by default it’s hidden from screen sharing. Closing the window doesn’t stop live transcription. The draft isn’t saved to the archive — the full transcript appears after the call as before. It needs the Parakeet model downloaded; while live transcription runs, the app keeps a second copy of the model in memory — if your Mac slows down during a video call, stop it.
Selecting several replies
The meeting transcript has a new “Выбрать несколько…” (Select several…) mode, also opened with the “Выбрать” (Select) button next to search: tick replies with the circle or by clicking an empty spot of a row, with ⇧ for a range. The bar at the bottom can reassign the selected replies to a participant or to “Новый участник…” (New participant…), copy them and delete them. Deleting and batch moving can’t be undone — check your selection first.
Long transcripts now scroll without stutter, and long monologues are split into paragraphs. Meetings and the main window open faster.
Calendar and day overview
A new “Календарь” (Calendar) item in the sidebar shows meetings on a week or month grid; click one to see its time, participants and summary. To jump to a meeting in the calendar, use “Показать в календаре” (Show in calendar) in its menu in the list.
While no meeting is selected, “Все встречи” (All meetings) shows an overview of the latest day: a timeline of meetings by hour, the day’s people, “Итоги дня” (Day summary), a question box across all meetings and a “Последние 30 дней” (Last 30 days) calendar — click a day to switch the overview to it. The day summary needs a connected language model; it’s built by itself only for the latest day and only with “Settings → Обработка → Запускать автоматически после распознавания” (Processing → Run automatically after recognition) on — otherwise use the “Собрать” (Build) button.
Project scenarios
Your own prompt scenarios can now run on a project too: create a scenario of the “По проекту” (By project) kind in “Settings → Сценарии” (Scenarios) and run it from the “Сценарии” (Scenarios) menu at the top of the project card. It takes into account the project’s meetings, including subprojects, and the project description; the result is kept in the project card, and the assistant sees it in that project’s chat.
Google Meet
The browser extension that shows who is speaking now supports Google Meet on a par with Kontur.Talk. The “Установить” (Install) button in “Settings → Кто говорит” (Who is speaking) leads to the Chrome Web Store.
Recording and stability
Fixed the cause of call recordings sometimes coming out empty. And if writing to disk still fails, VoiceGrep shows a warning with a sound, and the meeting isn’t deleted as empty.
Switching microphones mid-call — for example, connecting or removing a Bluetooth headset — no longer cuts off your track for the rest of the recording or splits the call into two meetings, and your replies are no longer shifted in time relative to the other people.
When auto-recording calls in Kontur.Talk, other Electron apps and Chromium-based browsers, only the call app’s audio is now recorded rather than all system audio (with “Писать только звук приложения звонка” (Record only the call app’s audio) on — the default). When all system audio is recorded — a manual recording, Safari — apps from the “Приложения, которые слушают микрофон” (Apps that listen to the microphone) list are no longer captured: a virtual microphone or noise suppressor no longer turns your voice into a second speaker. If a call in such an app starts after you press “Начать запись” (Start recording), its audio won’t be recorded.
Fixed a crash on macOS 15 when an open meeting card was deleted automatically — for example, as an empty recording.
Look and feel
Participant colors in a meeting card are easier to tell apart, section headings are larger, the app accent is teal, and the app has a new icon. An empty attachments section is hidden: attach a file with the paperclip in the meeting header or by dragging it onto the card. “Граф общения” (Social graph) has moved to the sidebar, under “Инструменты” (Tools). Player timecodes no longer stack into a column in meetings with many participants.
Version 1.12.0
New speech recognition model
Meetings are now transcribed with the Parakeet Ultra model. It’s noticeably more accurate than the previous one — it drops and mixes up words less often — while speed and load on your computer stay the same.
After the update, the new model downloads itself once — about 550 MB. While it’s downloading, meetings are transcribed with the previous model, so there’s nothing to wait for. The download doesn’t start during a call. You can see which model is transcribing right now in “Settings → Транскрипция → Модель Parakeet” (Transcription → Parakeet model). On a metered connection you can turn off the automatic download in the same place — then the model downloads on the first transcription or when you press the button.
Service tokens like “<unk>” and English “mm-hmm” on pauses are also gone from transcripts.
A new “Участники” (Participants) block
Instead of several cards above the transcript, there’s now a single block. Named people are shown as chips, and voices awaiting an answer as short questions: an unnamed “Speaker”, a voice the calendar or Talk calls by a different name, a person created from Talk data. Each question has a hint, a “послушать” (listen) button, a one-click answer and “Отменить” (Undo) for a few seconds after you answer.
The previous participants view opens via the “Все участники” (All participants) link in the block header, and you can bring it back entirely in “Settings → Транскрипция → Отображение встречи” (Transcription → Meeting view).
Speaker recognition
You can label a person in several meetings at once. If you turn on “Settings → Транскрипция → Правка участников → Предлагать подписать человека в других встречах” (Transcription → Editing participants → Suggest labeling the person in other meetings), then when you name a “Speaker”, VoiceGrep looks for the same voice in already transcribed meetings from two weeks before and after and offers to label them too. Nothing changes until you click “Подписать” (Label), and the labels can be undone. The feature is new and off by default: before labeling, open the list (“Показать”, Show) and uncheck the ones that don’t belong.
Recognition has become more careful: it recognizes familiar people more often and labels a voice with someone else’s name less often. A voice is remembered only once the person has spoken for at least 20 seconds. This means that after the update, labels on new meetings may differ slightly from what you’re used to — in borderline cases VoiceGrep now leaves “Speaker” more often rather than guessing.
If you talk without headphones and your voice comes back through the other person’s speakers, that echo is now cut out and no longer turns into their lines. And a phrase that used to be transcribed twice — for two different people — now stays with one.
Labels from Kontur.Talk
If you call in the Kontur.Talk app, VoiceGrep can read from it who is speaking right now and label participants more accurately: suggest names for unnamed voices, sort out lines when several people have merged into one voice, and create new people from their full names in Talk (such people are marked, and you confirm their name).
This is off by default. To turn it on: “Settings → Кто говорит → Читать, кто говорит, из приложения Толк” (Who’s speaking → Read who’s speaking from the Talk app), then “Разрешить…” (Allow…) — macOS will ask you to give VoiceGrep Accessibility access. Data from Talk never leaves your Mac.
People and the meeting view
“Граф общения” (Communication graph) opens in a separate window: groups of people, where time goes, who connects different teams and who you’re meeting less often.
New player: a speaker track right under the scrubber shows who spoke and when. Avatars got a thin ring in the person’s color.
“Открепить” (Detach) and “Удалить из встречи” (Remove from meeting) no longer delete a person who has voiceprints or tasks. VoiceGrep takes a database snapshot before merging people and removing a participant.
Recording and stability
The microphone no longer drops silently on a stereo input, an 8 kHz headset, a 96 kHz audio interface or inputs with three or more channels. Re-encoding such recordings no longer crashes the app.
The window no longer hangs after splitting and merging meetings, when checking recognition, or when dragging the column divider. Fixed a rare crash that could happen even when idle — at night, for example.
Live transcription during a call — a draft of the text right in the meeting view. It’s still experimental and off; turn it on in “Settings → Запись → Живая расшифровка во время звонка” (Recording → Live transcription during a call).
Other
External access (MCP): when many apps are connected to the archive at once, new ones are no longer refused — idle connections free up their slot and reconnect on their own with the next request.
Version 1.11.0
Calendar link via OWA Widget
If you have OWA Widget installed (version 1.0.55 or newer), VoiceGrep can figure out on its own which calendar meeting was recorded and take its title, invitee list and agenda from there. A meeting is linked automatically only when the answer is unambiguous. If several meetings were happening at that time, the meeting view will ask “Какая это была встреча?” (Which meeting was this?).
The calendar meeting’s subject is shown in the meeting view header. Invitees and the agenda help the summary, tasks and the assistant. If the recording has unnamed voices and the invite lists people who aren’t among the participants yet, the meeting view will suggest their names.
This is off by default. Turn it on in “Settings → Данные → Календарь” (Data → Calendar); OWA Widget itself must have AI access enabled (“Settings → AI (MCP)” in OWA Widget). The first time VoiceGrep connects, OWA Widget will ask whether to let it in — click “Разрешить” (Allow).
External access (MCP): permission for each app
If you have “Доступ извне (MCP)” (External access) turned on, VoiceGrep now asks for permission for each app that connects to the archive. After the update, the “Разрешить / Отклонить” (Allow / Deny) prompt will also appear once for apps that were already connected: the window pops up at the top center of the screen the first time they make a request. The list of apps and the “Отозвать” (Revoke) button are in “Settings → Данные → Доступ извне (MCP) → Клиенты” (Data → External access → Clients); the check can also be turned off there.
Connected apps can also now see an ongoing recording (when it started and in which app, no text) and a meeting’s place in the transcription queue, and they find people by name in any word order: “Vlad Karetko” finds “Karetko Vlad”.
Other
The license is now verified more reliably. If “Settings → Лицензия” (License) says the copy of the app has been modified, download VoiceGrep again from the website and install it over the old one: your meeting archive won’t be affected.
Rows in the meeting list are now single-line, which makes the list more stable.
Version 1.10.0
VoiceGrep is now notarized by Apple
The app is signed with an Apple Developer certificate and has passed Apple’s check (notarization). New users no longer need to lift the block through Terminal: VoiceGrep opens like any other app.
Once, after this update
Because of the new signature, macOS treats VoiceGrep as a new app, so after updating you need to grant access again, once. Your recordings, transcripts and settings stay where they are.
Keychain. On first launch, macOS may ask whether VoiceGrep may use data from your keychain. Enter your Mac password and click “Always Allow”. That’s where the license and the language model key are stored: if you deny it, VoiceGrep won’t see the license and won’t start recording. If you did deny it, restart VoiceGrep and the prompt will appear again.
Screen Recording and Microphone. Open “VoiceGrep Settings → Разрешения” (Permissions) and grant both again. If VoiceGrep is already checked in “System Settings → Privacy & Security → Screen Recording” but the app still asks for access, remove the old entry with the “−” button and add it again. After granting screen recording access, restart VoiceGrep.
Launch at login. If VoiceGrep used to start when you logged in, check the toggle “Settings → Запись (Recording) → Автостарт (Autostart) → Запускать при входе в систему” (Launch at login).
Other
The link in “Settings → О приложении” (About) now leads to the new website voicegrep.com.
Version 1.9.0
Meeting links
Every meeting now has a link that opens it directly in VoiceGrep. Right-click a meeting in the list → “Скопировать ссылку на встречу” (Copy meeting link) and paste it into a note, task or email: clicking the link opens the meeting.
If your archive is connected to Claude Desktop, Claude Code or other MCP clients, the client itself gives links to the meetings it talks about. Clicking one in the chat opens a short page in the browser, which then opens the meeting in VoiceGrep. The browser may ask once for permission to open VoiceGrep — allow it.
External AI clients (MCP): meeting titles and projects
Clients can now rename meetings and create projects (including nested ones), as well as rename and delete them. For example: “title this week’s meetings by topic” or “create a Q4 project and put the meetings there”. Deleting a project keeps its meetings — only their link to the project goes away. A client can’t delete a project that has your note, materials or an assistant chat — that’s done in the app.
This is off by default and has its own toggle, “Разрешить переименовывать встречи и управлять проектами” (Allow renaming meetings and managing projects), in “Settings → Данные → Доступ извне (MCP)” (Data → External access). An already connected client will see it after reconnecting. Every change is logged to the mcp.log file next to the database.
Fixes
A project’s view no longer brings back the old name if the project was renamed while the view was open.
Version 1.8.0
Search finds words by how they sound
The assistant and archive search now compare not just spelling but also how words sound. If the transcription spelled a term the way it heard it, you can still find it: a query for “QMS” finds the Russian word “kumys”. Exact matches still come first, and the assistant tells verbatim mentions apart from sound-alikes. You can also search meeting titles and their summaries.
Search no longer freezes the window on a large archive: the index is built in the background.
External AI clients (MCP): filing into projects and stable connections
If your archive is connected to Claude Desktop, Claude Code, Cursor or other clients, you can ask them to file meetings into projects — for example, “sort this week’s meetings into projects”. The client adds meetings to existing projects and removes them from there. It can’t create, rename or delete projects, or change the meetings themselves. This is off by default and has its own toggle in “Settings → Данные → Доступ извне (MCP)” (Data → External access). An already connected client will see it after reconnecting.
Clients no longer ask for confirmation on every archive request: reads are now correctly marked as reads. Up to 64 clients can connect at once: previously a few open but idle sessions could take up every slot, and a new client was refused. The same settings show how many clients are connected right now.
Database backups no longer fill up the disk
The app made a copy of the database on every launch, and over time the backups folder took up more and more disk space. Now a copy is made at most once every six hours, and the app keeps the three most recent, one for each of the last seven days, and no more than 3 GB in total. On the first launch of the new version, extra old copies are deleted. Your meetings, recordings and settings are not affected.
Version 1.7.0
Process several meetings at once
Select any number of meetings in the list and right-click: actions that used to require opening each meeting one by one are now right there — “В избранное” (Add to favorites), “Добавить в проект” (Add to project) and “Обработать через LLM” (Process with LLM). The last one is how you catch up on an older archive: a single pass over all selected meetings in a row — “Улучшить текст” (Improve text), “Собрать саммари” (Build summary), “Извлечь задачи” (Extract tasks) or “Придумать название” (Suggest a title).
You choose the pass — the toggles in the processing settings don’t affect a batch. Before starting, the app tells you how much it’s about to process and asks for confirmation: every meeting uses requests to your model. Meetings are then processed one after another; the list shows “Обработка через LLM” (Processing with LLM) and “В очереди” (Queued), and a bar in the current meeting’s view shows the position (“3 из 12”, i.e. 3 of 12). You can stop at any time — right-click the selection, or use the ✨ menu in the meeting view; work already done is kept.
Meetings without a finished transcript are skipped, and if not the whole selection was processed, the app tells you at the end. In a batch, “Придумать название” (Suggest a title) leaves alone meetings you’ve named yourself. The item appears once the language model is enabled in Settings.
Bookmarks: full text, editing and jumping to the moment
The bookmarks you add during a call are now gathered in the meeting view as a single timeline: type icon, timecode and the full text of your note — it’s no longer cut off. Right-click to copy a note, edit it or add one to a bookmark set from the menu bar without a comment, or delete the bookmark. A button in the section header copies all of the meeting’s bookmarks at once as a list with timecodes.
Once a meeting is processed, clicking a bookmark starts playback ten seconds before the marked moment and scrolls the transcript to the same spot.
While recording, the same block is available right in the meeting view — with a timer, four bookmark buttons and a “заметка к моменту…” (note for this moment…) field. It’s there for when the floating panel is turned off or gets in the way. The assistant in the meeting view also sees your bookmarks with timecodes, so you can ask “what did I bookmark in this meeting?”.
Version 1.6.0
Materials: documents and notes alongside the transcript
You can now attach files to a meeting or a project — an agenda, slides from the call, a contract, a screenshot. The app keeps its own copy, so renaming or deleting the original on disk won’t break anything. PDF, Word, presentations and spreadsheets are converted to text, and the assistant takes them into account when answering: “what does the contract they sent say about deadlines?” becomes a question for the meeting, not for a file you open separately.
You can also write a note right in the app — it’s a material like any other, just typed by you, and unlike an attached document there’s no second copy of it anywhere, so notes are included in the regular database snapshot. When meetings are merged, materials move to the combined meeting; when a meeting is split, they stay with the first one.
If the app crashes when opening a meeting with an attached document, text extraction can be turned off: Settings → Данные → Материалы → “Извлекать текст из документов” (Data → Materials → Extract text from documents). Text that’s already been extracted is kept, and turning the option back on resumes extraction.
Open your archive to an external assistant
VoiceGrep can expose its archive to Claude Desktop, Claude Code, Codex or Cursor: the same transcript search, meeting reading, participants, tasks and materials, but from the work tool you already use. Everything stays on your machine, nothing goes out to the internet, and access is read-only — nothing can be changed from outside.
Off by default. Turn it on in Settings → “Доступ извне (MCP)” (External access), where you’ll also find ready-made config snippets for each client — just copy and paste.
A voiceprint directory for teams
When the same group of people has calls with several colleagues, recognition no longer starts from scratch for each of them: a shared directory lets you recognize voices that someone else has already named. The directory only suggests — it never renames anyone on its own and doesn’t interfere with recognition; the decision is always yours.
Off by default; turn it on in Settings → “Каталог команды” (Team directory). Keep in mind that voiceprints are biometric data: share them only within your team.
The “Знания” (Knowledge) section has been removed
The experiment with a fact base that the assistant maintained on its own didn’t pay off, and the section has been removed entirely — along with the two toggles for automatic fact capture in the processing settings. Accumulated facts will be deleted the first time this version launches, and they can’t be recovered. Meetings, transcripts, people, tasks, projects and voiceprints are not affected — only the contents of the “Знания” (Knowledge) section itself go away.
Faster and more reliable
The meeting card no longer hangs the app. On long transcripts the app could get stuck redrawing endlessly: the window froze and the fans roared while audio kept playing. Line text now uses its own rendering — selection and copying work as before.
The window opens right away. Clicking the Dock icon after a long idle period (when another app is active during a call) brought up the window with a noticeable delay — at that moment the app was re-reading the entire archive. Now it shows the window first.
“Проверка узнавания” (Recognition check) no longer computes the same thing twice — the badge on the card and the list itself now share a single result, and edits inside the list apply noticeably faster.
Two copies of the app no longer interfere with each other. If a second copy happened to be running, it could mistake the other copy’s live recording for an abandoned one and start transcribing it mid-call — leaving a transcript in the archive without a single line of yours. Now a live recording is visible as such, and the second copy leaves the other’s work alone.
Version 1.5.0
The transcription model downloads itself — and no longer depends on a VPN
The model used to be downloaded from HuggingFace, which some users simply can’t reach: without a VPN the app was useless. Now the model comes from our own server, and if that server doesn’t respond for some reason, the app quietly falls back to the old route — so things can’t get worse than before.
The download starts at launch, not at the end of your first call: by the time the first transcript appears, the model is usually already in place. Progress is shown as a line in the VoiceGrep menu in the menu bar. A dropped connection no longer starts everything over — the download resumes where it left off. Automatic download can be turned off in Settings → Распознавание (Recognition).
The models for telling voices apart now simply ship inside the app — that download is gone entirely, and your first meeting no longer waits on the network. And if there’s no network at all, the transcription model can be installed from a file downloaded in advance: Settings → Распознавание → “Установить из файла…” (Recognition → Install from file…). The same place now has “Удалить модель” (Delete model) — to free up space or reinstall a damaged copy.
Recording the other side no longer cuts out when the screen goes dark
An unpleasant quirk of macOS: when the display turns off, the system also turns off system-audio capture. On an eighteen-minute call the screen went dark from inactivity at minute eight — from then on only your microphone was recorded, and ten minutes of the other participants’ speech were lost. From the outside the recording looked normal.
There are now two safeguards. While recording, the display doesn’t sleep from inactivity — the “Не давать экрану гаснуть во время записи” (Keep the display on while recording) toggle in Settings → Запись (Recording). Closing the lid and putting the Mac to sleep manually work as usual. And if recording of the other participants does get interrupted, VoiceGrep restarts it on its own and keeps writing to the same file, filling the gap with silence so the tracks stay in sync.
If a gap does happen, the meeting gets a warning, “Дорожка собеседников записалась не полностью” (The other participants’ track wasn’t fully recorded), saying how much time was lost — the failure is now visible instead of being discovered through a strange transcript. The flip side of the first safeguard: a recording you forgot about used to end when the computer went to sleep; now only the time-based auto-stops (in the same recording settings) will end it — and until then the screen stays on.
A fresh install is ready to go without a trip through Settings
On new installs, automatic call recording, summaries and tasks are on from the start. The first-run setup assistant now opens by itself (previously the app silently lived as a single icon in the menu bar, and nobody granted it permissions) and has a new step for auto-recording and launching at login. Existing VoiceGrep users are not affected: your settings stay exactly as they were.
Miscellaneous
An empty meeting is no longer deleted if capture was failing. The app removes recordings in which not a single word was spoken — but if the sound was dropping out at that moment because of the capture itself, the silence may not be yours. Such a meeting is now kept, with a warning: a forty-minute call whose capture died at the start no longer disappears along with the only trace of the failure.
“Пересобрать саммари” (Rebuild summary) and “Распознать заново” (Re-transcribe) no longer get in each other’s way. Run back to back, they could write a summary on top of a transcript that no longer existed. Now the second action waits for the first, and an unavailable button explains why it’s unavailable — before you confirm replacing the transcript, not after.
New language model providers: DeepSeek and Polza AI — in the preset list under Settings → Обработка (Processing).
Version 1.4.0
Your own recordings — now in VoiceGrep too
You can create a meeting from an existing file: “Файл → Импортировать запись…” (File → Import recording…, ⌘I), or the same item in the VoiceGrep menu in the menu bar — it works even with the window closed. Both audio and video work (only the sound is taken from video), and the file is copied into the archive, so you can delete the original afterwards. Transcription starts automatically; if a call is being recorded at that moment, the import waits for it to finish.
In such a recording all voices are on a single track, so the app doesn’t know which one is yours — it asks after transcription, with a single “which one is me” question. Until you answer, no summary or tasks are generated: without “me” they would end up being about a stranger.
The app will refuse to import a file with no sound, and it will notice and warn you if you import the same file again — in both cases you can insist and continue.
Transcript text correction is now a separate step — and off by default
Previously the language model rewrote every line after each transcription. Measured across the archive, this step took 82% of all processing time — three times as long as the transcription itself — while changing about 15% of the words. Summaries, meeting titles, tasks and memory work just as well without it, so it’s now off by default: meetings finish processing noticeably faster.
If you want text correction, turn it on in Settings → Обработка → Править текст расшифровки (Processing → Edit transcript text). For an individual meeting you can run it by hand at any time: the ✨ menu at the top of the card → “Улучшить текст” (Improve text). Corrections already made stay as they are.
Select text across paragraphs
Meeting summaries, scenario results, person profiles, project overviews and finished assistant answers used to be selectable one paragraph at a time: you couldn’t drag a selection across several paragraphs or list items. Now you can — and copy it with ⌘C, list markers included. To copy an entire assistant answer, use the button below it (it appears on hover) or “Копировать ответ” (Copy answer) in the right-click menu.
Two small things worth knowing. If a long text is clipped in height, a selection dragged past the visible edge also picks up the hidden part — if you want to see what you’re copying, click “Показать полностью” (Show all) first. And clicking the text of an answer in the chat moves the input cursor from the question field into the answer itself: to keep typing, click the field.
Data protection
Two copies of the app on one store no longer stay silent. If VoiceGrep is already running, a second copy says so at launch instead of quietly running alongside it. And if the app fails to save data, it tells you right away rather than after the fact. Recording doesn’t stop in that case — audio is written to disk either way.
Merging and splitting meetings no longer bloats the audio. These operations re-encode the tracks, and the results used to be four times the size of the originals (53 MB instead of 13 for a half-hour conversation). Now the size is preserved. Also, if the disk runs out of space while exporting audio, the app reports an error instead of silently giving up.
Version 1.3.0
Tasks: statuses, due dates, priority and a board
The “Задачи” (Tasks) section is now complete: “Мои дела” (My tasks), “Жду от других” (Waiting on others), “Все задачи” (All tasks), “Доска” (Board) and “На разбор” (To review). A task now has four states — “Ожидает” (Pending), “В работе” (In progress), “Выполнена” (Done) and “Неактуальна” (No longer relevant), the last being a way to close out an agreement without deleting it. Change the status by swiping a row to the right, with a right-click, or in the task card.
The board has three columns, and cards move by drag and drop. At the top: a “Мои / Жду / Все” (Mine / Waiting / All) switch, search and a project filter (you can tick several at once, and “Без проекта” (No project) shows tasks outside any project). The “Выполнено” (Done) column keeps only what was closed in the last two weeks, so it doesn’t keep growing.
New: due dates (if a deadline was mentioned in the conversation, it’s stored both in words and as a date, so you can see where it came from; overdue tasks are highlighted), a “Важная” (Important) flag, picking a project separately from the meeting, and bulk actions on several selected tasks. You can also create a task without a meeting — ⌘N in the “Задачи” (Tasks) section.
Nested projects
A project can now live inside another — a tree up to three levels deep (“Clients” → “Acme” → “Rollout”). Drag a row onto the middle of another project to nest it, or onto the edge to reorder the list; while you drag, you can see what will happen. Project order is now yours, too, rather than alphabetical.
By default a parent project shows both its own meetings and its subprojects’ meetings — which is why the number next to the folder is larger than the number of meetings in it. The “Все / Свои” (All / Own) switch in the list header leaves only its own. The meeting itself stays in exactly the project you put it in. Deleting a project that has subprojects asks whether to move them up or delete them too; meetings are never deleted.
Recording
No more stray recordings from apps that listen to the microphone. Dictation, noise suppression, voice assistants — to VoiceGrep all of these looked like the start of a call, and auto-recording kicked in again and again. Settings → Запись → “Приложения, которые слушают микрофон” (Recording → Apps that listen to the microphone) now lists all such apps (the list fills itself). Tick the ones that shouldn’t count as a call — recording won’t start, and one already running will stop. If any other app takes the microphone alongside a ticked one, it counts as a call again: a real conversation won’t be lost.
Bookmarks and screenshots — now from the menu bar too. While recording, the VoiceGrep menu gets a “Во время записи” (During recording) section: the same four bookmarks (“Важно”, “Вопрос”, “Задача”, “Решение” — Important, Question, Task, Decision) and a screenshot (one item per display if you have several). A fallback for when the floating panel is turned off or in the way.
See which app a call was recorded from. The meeting card header shows the app the conversation happened in (“Talk”, “Zoom”, “Telegram”) next to the date and duration. It’s detected when recording starts; if it can’t be detected, there’s simply no label — the app doesn’t guess (this happens with calls in a browser).
Meetings
Split one meeting into two. If a recording captured two calls back to back, you can cut it by hand: “···” above the transcript → “Разделить встречу…” (Split meeting…). Possible boundaries are listed by time with the length of the pause, and the most likely one is suggested right away; for the selected one you see the seam — the last line of the first meeting and the first line of the second, with playback. Lines are never cut in half. Cut in the wrong place? Just merge them back.
Export the audio as a single file. “···” → “Выгрузить аудиозапись…” (Export audio…), or the row’s context menu in the list. Your microphone and the other participants’ audio are mixed into one .m4a that sounds the way the conversation went. Progress is shown, and the mixdown can be cancelled.
Share of participation. Above the “Участники” (Participants) card there’s a compact bar showing what share of the words each person spoke. It counts words, not time (the durations of individual lines are unreliable). Turn it off in Settings → Транскрипция → Отображение встречи (Transcription → Meeting display).
The buttons at the top of the card, tidied up. Language-model actions are now gathered in a single ✨ menu — “Улучшить текст” (Improve text), “Извлечь задачи” (Extract tasks), “Придумать название” (Suggest a title) and your scenarios; previously they were separate buttons. “Распознать заново” (Re-transcribe) now asks for confirmation: this run replaces the transcript, and speaker edits, line splits and model improvements will be lost. And if the new run doesn’t recognize a single line (say, the model failed to load), the previous transcript stays in place.
Speaker recognition
Fewer stray “Speakers”. If there’s only one other person on the call, short fragments of their speech — interruptions, “uh-huh,” “I mean” — are attributed to them instead of creating a separate unnamed participant. When there are several voices, the app doesn’t do this.
Two familiar people no longer get merged into one voice. If the labeling lumped two people’s speech into one cluster, the app now more often splits them back apart — including when only one of the two is recognized.
Showing and learning are now separate. Names are filled in as before, but the app is now more careful about adding to a voiceprint — only when the match is confident. This cures the slow “contamination” of voiceprints that made recognition degrade over time. The change applies only to new transcriptions; meetings already processed aren’t recalculated.
Empty participants are removed automatically. If, after your edit, an unnamed “Speaker” has no lines of their own left — only short interjections inside other people’s blocks — it’s removed from the meeting automatically, with a notice at the top. This never applies to named people or to you. Turn it off in Settings → Транскрипция → Правка участников (Transcription → Participant editing); there’s also a “Убрать пустых (N)” (Remove empty (N)) button above the participant list.
When you reassign a line and there are many participants, the menu now has a short “Похоже на голос” (Sounds like) section at the top — the two or three people whose voices are closest in this meeting. It’s a hint, not a decision: the full alphabetical list is still below.
Fixes
The app crashed if you deleted a line while the background recognition check was working through the archive. If that coincided with a live call, the recording was cut off. Now the background check stops and starts over instead of crashing the app.
Version 1.2.0
Meeting titles instead of “Meeting on July 14”
You can now name a call while it’s still recording: the floating panel has a new field for the title. It always works and needs nothing else.
The rest is for those with a language model connected (Settings → “Языковая модель” (Language model) → “Улучшать транскрипт через LLM” (Improve transcript with LLM)). The meeting card has a new action, “Придумать название” (Suggest a title): one click, and the model reads the transcript and proposes a title. If no model is connected, the app tells you so.
This can also happen on its own if, under Settings → “Обработка” (Processing), both “Запускать автоматически после распознавания” (Run automatically after transcription) and “Предлагать название встречи” (Suggest meeting title) are on (the second is on by default). A fresh meeting then gets a title right after transcription. The automation never touches meetings you renamed by hand — your title always wins. The prompt sent to the model can be edited in the same place.
Speaker recognition
“Запомнить голос” (Remember voice). The app used to promise it would learn a voice on its own — but the first voiceprint never appeared: to capture one, the person had to be recognized first, and recognizing them without a voiceprint is impossible. It was a closed loop, and the same voice kept spawning new unnamed “Speakers” — even for people with hundreds of lines in the archive. Now the first voiceprint is captured with a button, in the recognition check panel and on the person’s card. It’s built from their speech in meetings that are already transcribed — the person needs to have spoken for at least half a minute in one of them, otherwise the voiceprint would be noisy and attract other people’s voices; in that case the button explains what’s missing.
“Этого участника не было — раскидать по голосу” (This participant wasn’t here — redistribute by voice). If the app attributed someone to a meeting who wasn’t there, their lines can be handed out to the real participants — each line goes to the closest voice. Before anything is applied you see a list where any row can be corrected by hand. If an unnamed “Speaker” is left without a single line of their own afterwards, the app offers to remove it.
The recognition check no longer invents problems. It used to report “broken labeling” where there was none: the conclusion rested on a couple of seconds of interjections. Now it draws conclusions only where there’s enough speech, and “we didn’t check this” is no longer passed off as “the model can’t tell the voices apart.”
Recording
There’s a new toggle, “Спрашивать подтверждение перед началом записи” (Ask for confirmation before recording) — Settings → Запись (Recording) → “Параметры автозапуска” (Auto-start options); the section is active only when auto-start itself is on. It’s off by default: as before, recording starts silently.
When it’s on, VoiceGrep asks first: while the call is going, a panel floats above your windows offering “Записывать” (Record) / “Не записывать” (Don’t record). Without an answer, nothing is recorded and nothing is written to disk. The panel stays on screen for the whole call — you can say yes midway, and recording starts from that moment. The cost: the start of the call is lost, about three to six seconds.
Fixes
Deleting a meeting now stops its background processing. Previously transcription kept running on the deleted meeting: a spinner hung around, the app wouldn’t quit with ⌘Q, and sometimes it crashed.
Release notes before version 1.2.0 were not published.
Recording needs an activation key: send the ID машины (Machine ID) from Settings via Telegram and we’ll issue one. About the key