Parrot 0.16.0 — it knows who said what

Parrot now knows who said what.

Until today, everyone on the other side of a call was one blob called "Them". This release is the biggest change since the copilot: Parrot tells the people on your call apart, entirely on your Mac.

Why this matters: speaker detection is the most requested feature in this category, and the apps that have it charge for it. Meetily puts diarization behind a $10/month Pro plan. Anarlog puts speaker identification behind a $15/month plan and does it in the cloud. In Parrot it is free, on-device, and in the default install. Your audio never leaves your Mac, there is no account, and there is no meter.

What you get:

• Speaker detection, on your Mac — After each call, a small local model (a one-time 13 MB download) figures out who spoke when. Your own voice is always right because your mic is recorded separately. A 16-minute call takes about 6 seconds to process.
• Name each voice by listening — A card appears after multi-person calls: "Heard 2 people besides you, listen and name them." Play a couple of short clips of a voice, type the name, and every line from that voice updates across the transcript, reports, and coaching. Got one line wrong? Right-click it and move just that line.
• Old meetings work too — Recordings are kept, so the "Detect speakers" button on any past meeting labels it retroactively.
• Remember voices (opt-in) — Turn it on in Settings and naming someone saves their voiceprint locally. Next call with them, Parrot suggests "Sounds like Gürkan, confirm?" It never applies a name without your click, the voiceprints never leave your Mac, and you can forget any voice (or all of them) anytime.
• Reports get names — Post-call reports and coaching now say "Gürkan flagged the onboarding flow" instead of "Them said".

Free on update: if you are on 0.14.0 or later, Parrot updates itself to this version automatically within a day, or use Check Now in Settings. No reinstall, no charge, same as always.

Speaker detection uses pyannote-derived models (CC-BY-4.0) via FluidAudio (Apache-2.0). Thanks to the FluidInference folks for excellent open-source work.

What's Changed

• Speaker detection: real on-device diarization (phase 1) in #40
• Speaker naming: hear each voice, give it a name (phase 2) in #41
• Remember voices: opt-in cross-call speaker recognition (phase 3) in #42
• Dev builds launch again: make now signs Sparkle inside-out (rode in with #41)

Full Changelog: v0.15.0...v0.16.0

Parrot 0.15.0

The onboarding got rebuilt after a teardown of a competitor's first-run flow.

• A friendlier permissions screen. Each permission is now a big button that says what it does for you ("Help Parrot hear you", "Help Parrot hear your meeting") and flips to a green confirmation once granted. Same privacy story as always: audio only, never your screen, nothing leaves your Mac.
• A proper finish. The tour ends with a small celebration and a "Let's start" button.
• Replay it any time. Help → Show Welcome Tour, or Settings → General. Handy when permissions act up and you want the guided path again.
• Fixed: source builds crashed at launch. Since 0.14.0, make run produced an app that died instantly (the bundled Sparkle framework needed re-signing with your identity). Downloaded releases were never affected.
• User guide refreshed with the new screens and updated Settings wording.

What's Changed: new onboarding, re-runnable welcome tour, and the dev-build fix in #39

Full Changelog: v0.14.0...v0.15.0

Parrot 0.14.0 — the last update you install by hand

Parrot used to tell you a new version existed and then make you go and fetch it. From this release on, it just updates itself.

Self-updating. Parrot checks daily, downloads new versions in the background, and swaps itself out when you quit. Never during a recording, because the install only ever happens after the app closes. There's a Keep Parrot up to date toggle in Settings → General if you'd rather it asked, and Check Now is still there for the impatient.

This is the handover release. If you're on 0.13.0 or earlier, download this one the old way. It's the last time. Everything after it arrives on its own.

Updates are signed, and Parrot checks. Every update must carry a signature made with a key that exists only on my machine. If someone took over the update feed tomorrow, your copy of Parrot would refuse the download rather than install it. Nothing to set up on your end.

What's Changed

• Self-updating: replace the download banner with Sparkle in #37

Full Changelog: v0.13.0...v0.14.0

Parrot 0.13.0 — tell me what's broken

Parrot can now be told when it misbehaves without you leaving the app, the help is where you'd actually look for it, and the live transcript stopped fidgeting.

Report a bug without hunting for GitHub. There's a small ladybug in the bottom right corner of the window, and Help → Report a Bug… does the same thing. Write what happened, and Parrot fills in the boring parts: version, model, transcription settings, and optionally a picture of the window. Nothing is sent anywhere on its own. You get a pre-filled issue in your browser, you look it over, you post it. Ideas and feature requests go through the same button. The screenshot is a picture of Parrot's window only, never your whole screen, and it lands on your clipboard so you paste it in yourself.

The user guide is findable now. An "Open User Guide" link in Settings, a proper help button in the corner of the settings sidebar, and an About row that opens my hello page. Two things were quietly broken behind that: the in-app help was indexed in a format modern macOS can't read, and it's now built in both. Check Now tells you when you're already up to date instead of looking like a dead button.

The live transcript sits still. The typing bubble used to restart each sentence from the first word every couple of seconds while you spoke. It reads forward now, and the finished line lands sooner, which means Copilot cards do too.

Onboarding offers the compressed model. Large V3 Turbo Compressed (626 MB) has been in Settings since 0.12.1 but was missing from the first-run picker, where it's arguably the best choice: nearly the accuracy of the 1.6 GB model at a third of the download and memory. Model names read like names now, not like folder paths.

A filter I decided not to ship. Loud typing near the mic can make Whisper invent a confident sentence, and the obvious fix is to drop low-confidence lines. Measured against real transcripts, that would have been a disaster: Whisper's confidence tracks language, not nonsense. Segments in Turkish average -1.43 where English averages -0.29, so any threshold low enough to catch the junk would have silently deleted most of a bilingual user's transcript. The measurements are in the code so nobody tries it again, including me.

What's Changed

• Preview: stop restarting the typing bubble every cycle in #29
• Help discoverability: guide links in Settings, About row, modern help index in #34
• Bug reporting: corner button, report sheet, pre-filled GitHub issue in #35
• Onboarding offers the compressed model; confidence-floor filter investigated and rejected in #36

Full Changelog: v0.12.2...v0.13.0

Parrot 0.12.2 — watch your model download

Yesterday's 0.12.1 made the big models download; today they stop being mysterious about it. This release is carried by a community contribution from @jiqqaman.

You can see the download now. Picking a model shows a real progress bar with a percentage — in Settings, onboarding, and the dashboard — and it names the model it's fetching, so no more staring at a spinner wondering if it's stuck, or which of the five models it's even getting. When the download finishes and Core ML spends a minute preparing, it says that too. Built by @jiqqaman in #31.

Half-finished downloads fix themselves. If a download gets interrupted, Parrot used to mistake the leftover folder for an installed model and fail on load. It now verifies every required file and resumes what's missing instead. Also @jiqqaman.

Upgrades from 0.11.x keep their Turbo. The matcher now accepts a previously downloaded turbo folder instead of forcing the 1.6 GB re-download that 0.12.1 asked of everyone. If you already re-downloaded yesterday: sorry, that one's on me.

Switching models mid-download behaves. Starting a new download cancels the old one instead of racing it — previously the last one to finish won, which could load a model you'd already navigated away from.

What's Changed

• Model download progress, resume, and safer switching (adapts #31 by @jiqqaman) in #33

Thanks @jiqqaman for the contribution that carries this release, and @anibalsouza + @onraz for the reports on #30 that started it all.

Full Changelog: v0.12.1...v0.12.2

Parrot 0.12.1 — the big model actually downloads

A small release for a big model: Large V3 Turbo couldn't download on new machines. Now it can, and it brought a lighter sibling.

Large V3 Turbo installs again. Picking it on a machine that didn't already have it failed with "Model Not Found" (#30): a fresh download asked the model library for a folder name that doesn't exist there. Parrot now asks for the right one, so the download just works — nothing to configure. Thanks @anibalsouza for the report with the exact error in it.

New model: Large V3 Turbo Compressed. The same OpenAI turbo checkpoint at 626 MB — near-best accuracy with far less memory and disk, and a much lighter first download. Suggested by @onraz. It's in Settings → Transcription, and the user guide's settings page covers it.

A heads-up for existing Turbo users. The Turbo option now fetches OpenAI's genuine turbo checkpoint, so your first launch after updating re-downloads it once (1.6 GB — the old label said 1.5, it fibbed).

What's Changed

• Fix Large V3 Turbo model download, add 626 MB compressed variant in #32

Full Changelog: v0.12.0...v0.12.1

Parrot 0.12.0 — sentences, not fragments

The biggest update since Parrot went public, and the first one under GPL-3.0.

Transcripts read like people talk now. Lines land as whole sentences when the speaker pauses, instead of chopped 2-second fragments, and while someone is mid-sentence you see a live preview filling in. Silence is never transcribed anymore, which also gets rid of the phantom "you" and "Thank you." lines that used to appear out of nowhere.

The quit crash is gone. Quitting mid-recording used to leave the app unable to start again (#18). Quitting now finishes the recording properly first, and a recording interrupted by a crash or force-quit is rescued on the next launch. Thanks @Ennazk for the crash report that cracked it.

You control what the Copilot costs. New in Settings → Copilot: a pace picker (Fast, Balanced, Relaxed — Relaxed spaces requests out to fit free model tiers), a setting for how much conversation each request carries, and a pause button right on the call screen. While paused, nothing is sent and nothing is spent, and your cards survive. Thanks @datpro12345 for the push (#20).

Two transcription bugs fixed. With a custom vocabulary set, Whisper could silently swallow a whole sentence; and when call audio plays through speakers, your mic could hear it and write the other side's lines twice. Both found by testing on real calls, both fixed.

Parrot has a real user guide now. turantekin.github.io/Parrot/help — screenshots of every screen, a proper walkthrough of profiles and settings, and troubleshooting. The same pages ship inside the app under Help → Parrot Help, searchable and offline.

License. From this release Parrot is GPL-3.0. Releases up to 0.11.3 remain MIT.

What's Changed

• Fix #18: quit-while-recording crash loop, main-actor lifecycle + graceful quit in #22
• Transcript segmentation: cut at speech pauses instead of fixed 2s chunks in #23
• Copilot budget controls: pace presets, context window, in-meeting pause in #24
• Relicense to GPL-3.0 in #25
• Fix glossary-prompt utterance loss + speaker-bleed dedup + live-loop test harness in #26
• Rolling preview: live text while an utterance accumulates in #27
• User guide: docs/help pages, in-app Apple Help Book, and a hello page in #28

Full Changelog: v0.11.3...v0.12.0

Parrot 0.11.3 — your mic fights back (first community release)

This is Parrot's first release with community contributions — thank you @wkoszek for three merged PRs and the bug report that drove the headline fix. 🎉

• Your mic fights back — the bug where transcription died the moment you joined a call (#12): when another app grabs the microphone, macOS silently feeds Parrot pure silence. Parrot now detects it within seconds, shows "mic muted by another app — reclaiming" in the device bar, keeps trying to take the mic back, and recovers the moment it frees up. Everyone else's audio keeps recording throughout.
• No more eternal "Loading WhisperKit model…" — models already on disk now load directly (works offline too), and a genuine failure shows a real error instead of a spinner that never ends.
• A smarter copilot — live cards no longer repeat the same issue in different words: the model now flags its own re-emissions and Parrot drops them, with a safety net so a genuinely new card is never lost.
• Faster, steadier analysis calls — fixed a hidden inefficiency that made the first copilot pass after every launch pay a heavy one-time API cost.
• Build from source in one command — make run (by @wkoszek). The README now shows real screenshots, a "Why trust an app that hears your calls?" section, and there are new CONTRIBUTING and SECURITY docs.
• Developer builds stop nagging about updates (by @wkoszek), and audio diagnostics for bug reports are now actually capturable from the system log.

What's Changed

• Add Makefile for building and launching from source by @wkoszek in #9
• Add AGENTS.md (CLAUDE.md symlink) and FILEMAP.md by @wkoszek in #10
• Don't nag about updates on non-numeric versions + AUDIODBG diagnostics by @wkoszek in #11
• README refresh: real screenshots, make-based build docs, trust section, CONTRIBUTING + SECURITY by @turantekin in #17
• Mic-loss watchdog + auto-reclaim, model-load fix, capturable diagnostics by @turantekin in #19
• Card dedup: model-side supersedes verdict, corroborated client-side by @turantekin in #21

Full Changelog: v0.11.2...v0.11.3

Parrot 0.11.2 — reports you can read, watch, and trust

• Reports, redesigned — every report section is now a card with a tinted icon and a title you can actually read: orange means needs attention, green means positive or committed, blue means info. Mixed bullet styles from the models now render cleanly too.
• Watch your report arrive — no more "No report was generated" while it was still being written. The Report tab shows live progress ("Writing your report…", then "Analyzing your coaching report…"), and the sidebar spinner now covers the whole report phase.
• Faster local reports — smarter context sizing stops Ollama from reloading the model between the summary and coaching passes, so back-to-back report calls stay warm.
• Honest reports — live copilot insights are now clearly fenced off as notes, and commitments must be explicitly said in the transcript before they appear in a report. This fixes fabricated commitments traced in real-call fact-checks.

What's Changed

• Report experience: section cards, live progress states, faster + honest local reports by @turantekin in #8

Full Changelog: v0.11.1...v0.11.2

Parrot 0.11.1 — you always know your version, and updates find you

• Version, visible — bottom of the sidebar and Settings → General → About show exactly which Parrot you're running.
• Updates find you — the app checks GitHub once a day; a new version shows a quiet banner with a one-click DMG download (never during a recording). Parrot → Check for Updates… now genuinely checks and tells you when you're already current.

What's Changed

• Update awareness: visible version + daily GitHub release check with banner by @turantekin in #7

Full Changelog: v0.11.0...v0.11.1