Transcripted

Record, transcribe, and identify speakers — entirely on your Mac.

Every conversation stays on your device. No cloud. No subscription. No compromise.

Contents

• Why Transcripted
• Install
• How It Works
• Features
• Sample Output
• Comparison
• Privacy & Security
• Configuration
• Troubleshooting
• Architecture
• Roadmap
• Contributing
• Acknowledgements
• License

Why Transcripted

• 100% Private — Audio never leaves your Mac. All models run locally on the Neural Engine. No network requests, no telemetry, no analytics.
• Speaker Identification — Learns voices over time with persistent fingerprints. Knows who said what across meetings, even weeks apart.
• Zero Cost — Free forever. No subscriptions, no usage limits, no "free tier" with asterisks. MIT licensed.
• Works With Everything — Captures audio from Zoom, Google Meet, Teams, Webex, FaceTime, Loom — anything that plays through your speakers.
• Set It and Forget It — Lives in your menu bar. Press ⌘⇧R or let auto-detection start recording when a meeting begins. Transcripts land in ~/Documents/Transcripted/ as clean Markdown.

Install

Download

1. Download the .dmg from the latest release
2. Drag Transcripted to Applications
3. Launch and grant permissions when prompted (Microphone required, Screen Recording for system audio)
4. Models download automatically on first launch (~600 MB for Parakeet, ~2.5 GB if Qwen is enabled)

Build from Source

git clone https://github.com/r3dbars/transcripted.git
cd transcripted
open Transcripted.xcodeproj

1. Select your Development Team under Signing & Capabilities
2. Press ⌘R to build and run
3. Models download automatically from HuggingFace on first launch

Requirements: macOS 14.2+, Xcode 15+, Swift 5.9+

Permissions

Permission	Purpose	Required
Microphone	Capture your voice	Yes
Screen Recording	Capture system audio from Zoom, Meet, Teams, etc.	For system audio
Accessibility	Global hotkey from other apps	For ⌘⇧R
Notifications	Transcript saved alerts	Optional

How It Works

Transcripted captures audio from your microphone and system output, runs speech-to-text locally on the Neural Engine, identifies who is speaking, and saves structured Markdown transcripts to your filesystem. Speaker fingerprints persist across sessions — the app learns voices over time.

Model	Role	Size
Parakeet TDT V3	Speech-to-text	~600 MB
Sortformer	Speaker diarization (up to 4 speakers)	Bundled
WeSpeaker	Voice embeddings (256-dim fingerprints)	Bundled
Qwen 3.5-4B	Speaker name inference from context	~2.5 GB (optional)

All models run locally via FluidAudio — no internet connection required after first download.

Features

Recording

• Floating pill UI — A Dynamic Island-style pill floats above all windows with aurora animations color-coded by audio source (coral for mic, teal for system audio). Minimal, beautiful, out of the way.
• Dual audio capture — Records both your microphone and system audio simultaneously. Hears both sides of every call.

• Auto meeting detection — Monitors for meeting apps and automatically starts recording when it detects sustained bidirectional speech. Stops when audio drops or the meeting app quits.
• Global hotkey — ⌘⇧R toggles recording from any app. No window switching needed.
• Recording health monitoring — Real-time quality tracking with capture quality grades, audio gap detection, and device switch monitoring.

Transcription

• Persistent voice fingerprints — Voice embeddings stored in a local SQLite database. Returning speakers are recognized across sessions, even weeks apart.
• Speaker name inference — On-device Qwen 3.5-4B LLM analyzes conversation context ("Hey Sarah, can you pull up the report?") to infer and assign speaker names automatically.
• Up to 4 speakers — Sortformer diarization handles multi-speaker conversations with speaker-turn detection.
• Smart post-processing — Automatic speaker merging, database-informed splitting, and name variant handling (Mike↔Michael, Nate↔Nathan, etc.).

Output

• Markdown transcripts — Clean .md files with YAML frontmatter (date, duration, word count, speaker metadata, capture quality metrics).
• [MM:SS] timestamps — Every utterance is timestamped for easy reference.
• Obsidian-ready — Optional Obsidian-compatible frontmatter with tags, aliases, and CSS classes.
• Auto-save — Transcripts saved to ~/Documents/Transcripted/ (customizable).
• Agent-ready — JSON sidecar files and index for automation workflows.

Stats

• Total hours transcribed, recording count, and active days
• Current and longest recording streaks
• Monthly activity heatmap
• Recent transcript quick-access

Sample Output

---
title: Weekly Standup
date: 2025-06-12T09:00:00
duration: 23m 47s
speakers: [Sarah Chen, Marcus Johnson, You]
words: 3842
capture_quality: excellent
audio_sources: [microphone, system_audio]
---

Sarah Chen [0:00]
Good morning everyone. Let's start with updates from the backend team.

Marcus Johnson [0:12]
Sure. We shipped the new caching layer yesterday. Response times dropped about 40%.

You [0:28]
Nice. I saw the metrics this morning — the P99 latency improvement is significant.

Comparison

vs. Cloud Transcription Services

	Transcripted	Otter.ai	Fireflies.ai	Fathom	tl;dv
Processing	100% local	Cloud	Cloud	Cloud	Cloud
Meeting bot joins call	No	Yes	Yes	Yes	Yes
Speaker diarization	✅	✅	✅	✅	✅
Learns voices over time	✅	❌	❌	❌	❌
Speaker name inference	✅ (on-device)	❌	❌	❌	❌
Works offline	✅	❌	❌	❌	❌
Open source	✅ MIT	❌	❌	❌	❌
Price	Free forever	$8–$30/mo	$10–$39/mo	$15–$39/mo	$18–$59/mo
Data location	Your Mac	US cloud	US cloud	US cloud	EU cloud
Audio retained	Deleted after transcription	Stored on servers	Stored on servers	Stored on servers	Stored on servers

vs. Privacy-Focused & Local Tools

	Transcripted	Granola	Krisp	MacWhisper	Meetily
Processing	100% local	Hybrid (cloud AI)	On-device	100% local	100% local
Speaker diarization	✅ Sortformer	❌ desktop	✅	⚠️ Beta	⚠️ PoC
Learns voices	✅	❌	❌	❌	❌
Name inference	✅ Qwen	❌	❌	❌	❌
Auto meeting detection	✅	❌	❌	❌	❌
Works fully offline	✅	❌	✅ (English)	✅	✅
Open source	✅ MIT	❌	❌	❌	✅ MIT
Cost	Free	$14–$35/mo	$8–$16/mo	$80 one-time	Free

vs. Apple Built-in

	Transcripted	Apple Dictation	Voice Memos	Notes (macOS 15+)
Speaker diarization	✅ up to 4	❌	❌	❌
System audio capture	✅	❌	❌	❌
Auto meeting detection	✅	❌	❌	❌
Voice fingerprints	✅	❌	❌	❌
Structured output	Markdown + YAML	Plain text	Searchable text	Inline text

Apple's SpeechAnalyzer API (WWDC 2025) adds speaker diarization but provides no persistent speaker ID, no meeting detection, and no system audio capture. Transcripted fills all three gaps.

Privacy & Security

• All processing is local. Audio is processed on your Mac's Neural Engine. No data is sent to any server.
• Audio is ephemeral. Raw audio is held in memory during recording and discarded after transcription.
• Transcripts are local files. Plain Markdown saved to your filesystem. You own them.
• No network access. No outbound connections during normal operation. Models are downloaded once on first launch.
• Open source. Every line of code is auditable.

See SECURITY.md for vulnerability reporting and the full privacy architecture.

Recording Ethics

• Know your local laws — many jurisdictions require consent from all parties.
• Inform participants when recording in professional settings.
• This tool is designed for your own productivity, not surveillance.

Configuration

All settings accessible via ⌘,:

Setting	Default	Description
Save location	~/Documents/Transcripted/	Where transcripts are saved
Your name	—	Used for speaker attribution
Auto-record meetings	Off	Automatically start when meeting apps detected
Qwen speaker naming	On	Use on-device LLM to infer speaker names
UI sounds	On	Play sounds on recording start/stop/save
Obsidian format	Off	Add Obsidian-compatible metadata
Aurora animation	Off	Enhanced recording animation

Keyboard Shortcuts: ⌘⇧R toggle recording  ·  ⌘, settings  ·  Escape dismiss transcript tray

Troubleshooting

No system audio captured

• Grant Screen Recording permission: System Settings → Privacy & Security → Screen Recording → enable Transcripted
• Restart Transcripted after granting the permission

Models not loading

• Check disk space (~3 GB needed)
• Check ~/Library/Logs/Transcripted/app.jsonl for detailed error logs
• Models download from HuggingFace on first launch — ensure internet for initial setup

Qwen not identifying speakers

• Ensure "Qwen speaker naming" is enabled in Settings
• Requires ~4 GB free RAM
• Works best with 15+ minutes of conversation containing name mentions

Transcript quality is poor

• Check capture_quality field in transcript frontmatter
• Ensure microphone is positioned correctly
• System audio quality depends on the meeting app's output

App doesn't appear in dock

• By design — Transcripted runs as a menu bar app (LSUIElement). Look for the icon in your menu bar.

Architecture

Project Structure

Transcripted/
├── Core/                        # Audio capture, transcription pipeline, persistence
│   ├── Audio.swift              # Microphone capture via AVAudioEngine
│   ├── SystemAudioCapture.swift # System audio via CoreAudio process taps
│   ├── Transcription.swift      # Orchestrates STT + diarization pipeline
│   ├── TranscriptionTaskManager.swift  # Background queue with retries
│   ├── TranscriptSaver.swift    # Markdown + YAML frontmatter output
│   ├── TranscriptStore.swift    # Transcript file discovery & parsing
│   ├── StatsService.swift       # Recording statistics aggregation
│   └── StatsDatabase.swift      # SQLite stats persistence
│
├── Services/                    # ML models & integrations
│   ├── ParakeetService.swift    # Parakeet TDT V3 (speech-to-text)
│   ├── SortformerService.swift  # Sortformer (speaker diarization)
│   ├── QwenService.swift        # Qwen 3.5-4B (speaker name inference)
│   ├── SpeakerDatabase.swift    # Voice fingerprint storage (SQLite)
│   ├── MeetingDetector.swift    # Auto-detection of meeting apps
│   └── EmbeddingClusterer.swift # Post-processing speaker segments
│
├── UI/
│   ├── FloatingPanel/           # Floating pill UI + aurora animations
│   └── Settings/                # Settings window + stats dashboard
│
├── Design/                      # Design tokens, colors, shared components
├── Onboarding/                  # First-run experience (5 steps)
└── TranscriptedApp.swift        # App entry point (menu bar app)

Threading Model

Component	Thread	Notes
Audio, Transcription, TaskManager	@MainActor	UI-bound state
PillStateManager, all Services	@MainActor	UI-bound state
SystemAudioCapture	DispatchQueue + NSLock	Real-time audio I/O
SpeakerDatabase, StatsDatabase	Serial DispatchQueue	Sync reads, async writes
CoreAudio I/O callbacks	Real-time thread	No I/O, locks, allocations, or ObjC calls

CoreAudio I/O callbacks run on real-time threads. Buffers are deep-copied before async dispatch — never processed in-place.

Data Storage

Data	Location	Format
Transcripts	~/Documents/Transcripted/	Markdown (.md)
Speaker database	~/Documents/Transcripted/speakers.sqlite	SQLite (WAL mode)
Recording stats	~/Documents/Transcripted/stats.sqlite	SQLite
Failed queue	~/Documents/Transcripted/failed_transcriptions.json	JSON
Speaker clips	~/Documents/Transcripted/speaker_clips/	WAV
Application logs	~/Library/Logs/Transcripted/app.jsonl	JSON Lines (rolling 2000)
Qwen model cache	~/Library/Caches/models/mlx-community/	MLX 4-bit quantized

All user data stays in ~/Documents/Transcripted/. No hidden directories, no cloud sync, no telemetry.

Speaker Identification Deep Dive

1. Embedding extraction — WeSpeaker generates a 256-dimensional, L2-normalized vector for each audio segment.

2. Adaptive matching — Cosine similarity against the speaker database with thresholds that relax as more segments are available:

Segments	Threshold	Rationale
1	0.85	High confidence with limited data
2–3	0.78	Slightly relaxed with more data
4+	0.70	Sufficient data for lower threshold

3. EMA blending — Matched embeddings are updated: new = 0.85 × old + 0.15 × new, allowing profiles to adapt over time.

4. Post-processing — Pairwise merge via union-find, database-informed splitting, and name variant groups (Mike↔Michael, Nate↔Nathan, etc.).

5. Name inference — Qwen 3.5-4B analyzes the first 15 minutes of transcript to infer names from conversational cues.

Meeting Auto-Detection

Monitored apps by bundle ID: Zoom (us.zoom.xos), Teams (com.microsoft.teams2), Webex (com.webex.meetingmanager), FaceTime (com.apple.FaceTime), Loom (com.loom.desktop).

Flow: App detected → lightweight audio metering → bidirectional speech for 5+ seconds → recording starts → audio drops for 15+ seconds or app quits → recording stops.

Roadmap

• Pre-built .dmg releases for one-click install
• Homebrew cask (brew install --cask transcripted)
• Real-time live transcription overlay
• Custom vocabulary / jargon support
• Multi-language transcription
• Transcript search across all recordings
• Calendar integration for meeting metadata
• Summary generation (key decisions, action items)

Have an idea? Open an issue.

Contributing

Contributions welcome — bug reports, feature requests, docs, or code. See CONTRIBUTING.md for setup and guidelines. Please read our Code of Conduct before contributing.

Acknowledgements

• NVIDIA NeMo — Parakeet TDT V3 and Sortformer models
• FluidAudio — Swift framework for running NeMo models on Apple Silicon
• WeSpeaker — Speaker embedding extraction
• Qwen — On-device LLM for speaker name inference
• MLX — Apple's machine learning framework for efficient on-device inference

License

MIT — use it, modify it, ship it.

---

If Transcripted is useful to you, a ⭐ helps others find it.