v1.1.18

What's New

• Auto-stop on silence — Recordings now stop automatically after a period of continuous silence, so a recording left running after a meeting ends (or a forgotten manual recording) won't keep going for hours. Configurable in Settings — Off / 5 / 10 / 15 / 30 minutes, defaulting to 5. Unlike the existing auto-record detection, this safety net applies to manual recordings too. Any audio resets the timer, so an active-but-quiet recording is never cut off early.

Install

Download NoteTaker-1.1.18.dmg, open it, and drag NoteTaker to your Applications folder.

v1.1.17

What's New

• Speaker attribution in summaries — When calendar participants are available, summaries now include a "Speaker Contributions" section that attributes key statements and contributions to individual speakers. The LLM uses contextual clues (self-references, being addressed by name, role-specific language) to identify who said what.
• Transcript post-processing — Filters out hallucinated and repetitive segments from transcripts, producing cleaner text before summarization.
• Improved content gating — Recordings shorter than 30 seconds are no longer sent to the LLM. Transcripts dominated by filler words (>50%) are also rejected, preventing low-quality summaries.

Install

Download NoteTaker-1.1.17.dmg, open it, and drag NoteTaker to your Applications folder.

v1.1.16 — Re-summarize with Different Models

What's New

• Re-summarize with different models — You can now re-summarize any existing meeting transcript using a different model. Open a meeting from History and click the "Re-summarize" button in the summary panel header. Switch models in Settings first, then re-summarize to compare quality across different LLMs.

Install

Download NoteTaker-1.1.16.dmg, open it, and drag NoteTaker to Applications. The app is signed and notarized by Apple.

v1.1.15

Expanded MLX Model Library

Added three new built-in models for local summarization:

Model	Size	Best For
Qwen3 30B MoE	~17 GB	Best summary quality. Recommended for 32GB RAM machines.
Qwen3 4B	~2.3 GB	Fast and capable. Recommended for 16GB RAM machines.
Llama 3.1 8B	~4 GB	Strong general-purpose middle ground.

Which should I choose?

• 32 GB RAM: Qwen3 30B MoE — best summaries, runs efficiently via mixture-of-experts
• 16 GB RAM: Qwen3 4B — excellent quality-to-size ratio
• Tight on RAM: Llama 3.2 3B or Qwen 2.5 3B (already built in)

All models run entirely on-device via MLX. No setup, no servers — just pick one in Settings and download.

v1.1.14

Fix: Transcript duplication and loss

• Fixed transcript duplication — Transcripts no longer contain cascading repeated text. The live transcript buffer was incorrectly committing the full accumulated text as each segment, causing massive duplication when segments were joined.
• Fixed transcript loss across SFSpeech resets — SFSpeech sometimes silently resets its recognition within a session (without error 209 or isFinal). Text from before the reset was being overwritten. Now detects within-session text shrinks and preserves the previous text before it's lost.
• Fixed model picker popup after transcription — Stopping a recording no longer triggers a WhisperKit model selection prompt. The streaming transcript path doesn't use WhisperKit, so the model requirement has been removed.

Full Changelog: v1.1.13...v1.1.14

v1.1.13

Custom System Prompt

• Editable system prompt — customise the instructions sent to the LLM when summarising meetings. Access via Settings → Summarization → Edit Prompt.
• Supports {{duration}}, {{context}}, and {{participants}} placeholders for dynamic values.
• Custom prompts persist across sessions.
• Reset to default at any time.

v1.1.12

What's New

• Improved meeting summaries — Richer structure: concise overview, key decisions with context, discussion highlights organized by topic, blockers, and concrete next steps
• Hallucination prevention — Short or noise-only recordings are no longer sent to the LLM for summarization
• Configurable mic volume — New mic gain slider (0.5x–5.0x) in Settings
• Voice processing removed — Fixed issue where AEC was making users inaudible to other meeting participants
• Custom MLX models — Add any HuggingFace MLX model by ID or URL, remove models you don't need
• Edit menu in Settings — Cmd+C/V/X/A keyboard shortcuts now work in text fields
• Better diagnostics — Detailed logging for auto-summarization decisions
• Backward-compatible — Existing meeting summaries display correctly with the new layout

Full Changelog: https://github.com/Incept5/note-taker/blob/main/CHANGELOG.md

NoteTaker v1.1.11

What's New

• Instant transcription — Transcripts are now available immediately when recording stops. The live SFSpeech text captured during recording is used directly, eliminating the long WhisperKit batch processing wait. WhisperKit is retained as an automatic fallback if SFSpeech was unavailable.
• No more truncated transcripts — Live text is accumulated into an append-only buffer of timestamped segments with session-reset detection. When SFSpeech's ~60-second recognition window expires, the current chunk is committed before a new session starts, so no text is ever lost.
• Mic echo cancellation — Enabled Apple's built-in voice processing (AEC) on the microphone input. Previously, the mic picked up speaker output that was already in the system audio capture, causing voices to appear twice in recordings. Ducking is disabled so system audio volume stays normal during recording.
• Version number in menu bar — The app version is now shown in the popover header next to the NoteTaker title.

Full Changelog: https://github.com/Incept5/note-taker/blob/main/CHANGELOG.md

---

Requirements: macOS 14.2+ (Sonoma), Apple Silicon (M1+), 16GB RAM

v1.1.10 — Live Transcription with Apple Speech Recognition

What's New

Real-time live transcript during recording — replaced the slow WhisperKit streaming (10-second chunked updates) with Apple's SFSpeechRecognizer for near-instant word-by-word transcription. Text appears as people speak, with a scrollable transcript area showing the latest ~15 lines.

WhisperKit batch transcription still runs after recording stops for the high-quality final stored transcript.

Changes

• Live speech recognition — SFSpeechRecognizer with automatic session restart, thread-safe buffer handling, and 50ms throttled UI updates
• Scrollable live transcript — flowing text with scrollbar to review earlier text during recording
• ~1.5GB less RAM while recording — WhisperKit no longer loads at recording start, only when batch transcription begins after stop
• On-device speech recognition toggle — new setting (defaults to on) keeps audio processing fully local
• Graceful degradation — if Speech Recognition permission is denied, recording works normally without live text

Bug Fix

• Fixed permission prompt loop when granting Screen Recording permission on fresh install — Speech Recognition authorization is now requested lazily on first recording, not at app launch

SHA-256

40c2b3ec9cba7a46e2dc6bf0e2b48c35d09f752b82100b0bd95b9497b99c1c0c

NoteTaker v1.1.9

What's New

• Calendar-driven auto-recording — NoteTaker can now auto-start recording when a calendar event begins. A new polling service checks EventKit and Google Calendar every 60 seconds for events starting within the next 2 minutes. Enable "Auto-record from calendar" in Settings. Off by default.
• Calendar end timer — Auto-stops recording 5 minutes after a calendar event's scheduled end time. Provides a reliable backstop for meeting end detection alongside silence monitoring.
• Improved silence detection — Replaced the strict consecutive-seconds approach with a rolling 45-second window. Auto-stops when 85%+ of samples are silent, so brief system sounds (notifications, alert tones) no longer reset detection.
• Keychain password prompts fixed — Switched Google Calendar token storage to the Data Protection keychain and cached tokens in memory. Eliminates repeated macOS Keychain password prompts on app rebuild or relaunch.
• Settings/History window fix — Fixed a bug where clicking Settings or History from the menu bar had no effect after the window had been previously opened and closed.

Full Changelog: v1.1.8...v1.1.9