MacParakeet 0.8.10

MacParakeet 0.8.10

Meetings recover more reliably, transcripts are easier to correct, and long meeting cleanup uses less memory.

• Choose the speaker count before transcribing a file or link. While speaker detection is on, Transcribe shows Speakers: Auto. Pick 1–12 for the next drop, file, or link, including every file in that batch. The choice resets to Auto afterward. A video already in the Library is transcribed again so the count applies. Meetings still use Re-transcribe for an exact count.
• Short replies stay with the speaker who said them. One-word turns such as “Yeah”, “Okay”, and “Yes” are no longer folded into a neighbor. Exact counts also handle long stretches of silence better, so a muted track is less likely to become extra speakers.
• Unclaimed words are not another person. In a diarized meeting, system-track words that no speaker claimed are labelled Unattributed, listed last, and left out of the speaker count. Re-transcribe an older meeting to see the new label. A meeting with speaker detection off still says Others.
• Search a transcript and replace the words. In a timed transcript, ⌘F opens find and ⌥⌘F opens find with replace. Replace changes the current match; Replace All changes every match. Each replacement is one undoable edit. Transcripts without timing still use the whole-transcript editor.
• A meeting microphone that fails to start is retried. If the microphone or system audio never arrives, the recording says so instead of continuing quietly. Switching audio devices while you are not recording no longer leaves the shared microphone stuck.
• Long meetings use less memory while they are running and while they are cleaned up. Live preview no longer holds the capture stream. If preview falls behind, some live words may skip, and the saved recording is kept. Cleaned-mic processing no longer holds the whole recording in memory. A one-hour synthetic render dropped from about 988 MiB to about 26 MiB of peak footprint. Raw audio is unchanged.
• Add a second dictation shortcut. Settings → Capture → Dictation → Additional shortcuts can keep Fn on the Mac keyboard and another key on an external keyboard active together. Both pairs start off until you set them. A modifier chord matches exactly: ⌘D no longer also starts dictation for ⇧⌘D, and the extra modifier still reaches the front app.
• Control a recording from a link, if you turn it on. Settings → Capture → Allow recording control from links accepts macparakeet://meeting/start, pause, resume, and stop. It is off until you enable it. Once enabled, other apps and websites can send those links.
• New libraries include meeting notes in Summary. Libraries you already have keep their current Summary setting. Turn notes off per prompt if you do not want them.
• Claude model lists match the current lineup, including Sonnet 5.5. Saved API keys stay put. Ordinary startup and settings checks no longer open Keychain password dialogs. If macOS blocks access to a saved key, AI Setup says how to allow it; the key is not removed.
• Stopping a transcription stays stopped. Cancelling during speaker detection or formatting no longer saves that run as a finished transcript. Deleting many Library rows no longer brings a deleted row back when the page refreshes.
• Long transcripts finish cleanup. AI cleanup no longer fails at 30 seconds. A long recording is cleaned in sections of at most 12,000 characters and joined.

If you use Anthropic or OpenRouter from the command line: a command that omits --model now uses Claude Sonnet 5.5. Pass --model to keep an older id.

Requires Apple Silicon and macOS 14.2 or later. Update with Check for Updates…, or download MacParakeet.dmg and drag it to Applications.

Coming from 0.8.8 or earlier? This update also includes everything in 0.8.9: saved API keys stay when AI is turned off, Apple Intelligence is for cleanup only, and AI Setup shows which provider each task will use.

Changes since 0.8.9

Bundled CLI 5.1.0

The app bundles CLI 5.1.0. Omitting --model on inline Anthropic and OpenRouter commands selects Claude Sonnet 5.5 (claude-sonnet-5-5 or anthropic/claude-sonnet-5.5). Explicit speaker-count transcription no longer over-splits digital silence. In a diarized meeting, unclaimed system-track words are labelled Unattributed and left out of speakerCount. Newly seeded Summary prompts include meeting notes; existing prompt settings are preserved. Keychain reads during CLI commands no longer open a password dialog. retranscribe --update returns the row the shared transcription service committed, so a concurrent edit or deletion is not overwritten. This build also waits 300 seconds for a non-streaming cloud response and cleans a long transcript in sections of at most 12,000 characters. See the CLI changelog. The standalone Homebrew CLI has a separate release channel.