Luxicon

On-device 1-on-1 recorder for managers. Record a sit-down with a direct report on your iPhone, get a speaker-labeled transcript that stays on your devices, and export it as clean markdown or JSON for the AI assistant of your choice — performance summaries, check-in prep, longitudinal review.

No cloud APIs. No per-minute pricing. No audio leaving the phone.

Luxicon is a free, open-source service of Davidson College, built by the college's Technology & Innovation team. If you like what you see, consider giving to Davidson.

How it works

1. Enroll your voice once (~15 seconds of reading aloud). Luxicon stores a 256-number voice fingerprint — not the audio.
2. Record a 1-on-1 with the phone on the table. When you stop, the app diarizes the conversation (who spoke when), transcribes each speaker turn, and matches your enrolled voice — so the transcript comes out labeled you and them automatically.
3. Export timestamped markdown (or structured JSON) with talk-time stats, ready to paste into Claude, ChatGPT, or your HR system's AI check-in notes.

All on-device, via soniqo/speech-swift and, on iOS 26+, Apple's system speech model:

• Diarization — Pyannote segmentation + WeSpeaker embeddings with constrained clustering, capped to 2 speakers for 1-on-1s
• Transcription — Apple's on-device system speech model (iOS 26+, default), or NVIDIA Parakeet TDT (CoreML) — the built-in engine and the pre-iOS-26 default; switchable in Settings
• Speaker ID — WeSpeaker enrollment matching (cosine similarity)

Models download on first use and are cached on-device: the diarization models (and, when the Parakeet engine is used, transcription) from Hugging Face — up to ~700 MB — plus a small live-caption model; Apple's system speech model is downloaded and managed by the OS itself. Summaries use the Apple Intelligence system model — OS-managed, no download — and require an iPhone 15 Pro or later on iOS 26 or later; on other devices, export the transcript and summarize it with any AI assistant.

Measured June 2026 on an M-series Mac (release build): a 50-second two-speaker meeting diarizes, transcribes, and speaker-matches in 4.6 s (0.09× real-time), with every turn correctly attributed. iPhones with recent A-series chips should land within a few multiples of that. Reproduce with luxicon-cli <any two-speaker wav> — it prints the real-time factor.

Structure

Sources/LuxiconKit/     Core pipeline (platform-neutral Swift package)
  MeetingPipeline.swift   diarize → per-turn ASR → speaker naming
  Models.swift            transcript, turns, stats, enrollment types
  Export.swift            markdown + JSON export
  MeetingSummarizer.swift Apple Intelligence summaries (FoundationModels)
  Vocabulary*.swift       user glossary + ASR correction pass
  LuxiconSync.swift       LAN sync protocol (TLS-PSK) + SyncPusher.swift
Sources/LuxiconCLI/     macOS command-line harness
Sources/LuxiconMCP/     MCP server + `listen` sync receiver
App/                    iOS app (SwiftUI, generated with xcodegen)
  Widgets/                Control Center control + Live Activity
Tests/                  swift-testing unit tests (offline, no models)
packaging/              Mac listener installer pkg (scripts/build-installer.sh)

Building

iOS app

Requires Xcode 26+ to build (the background-processing code uses iOS 26 SDK symbols, runtime-gated so it runs fine on iOS 18+ devices). Target device: iOS 18+, A13 or later recommended. The Swift package (CLI, MCP server) builds with Xcode 16+.

brew install xcodegen
cd App && xcodegen generate
open Luxicon.xcodeproj   # set your signing team, build & run on device

Note: diarization uses MLX (Metal) and does not run in the iOS Simulator. Use a physical device, or the CLI on a Mac.

macOS CLI

swift build
bash scripts/build_mlx_metallib.sh debug   # compile MLX Metal shaders
.build/debug/luxicon-cli meeting.wav \
    --enroll "Your Name=enrollment.wav" \
    --title "Weekly 1:1" --out ./out

First run may require the Metal toolchain: xcodebuild -downloadComponent MetalToolchain.

Other flags: --vocab "Choreo, OKR" / --vocab-file terms.json ground transcription in your jargon, and luxicon-cli push export.json --token <token> [--host <mac-ip>] exercises the same sync path the app uses.

MCP server (query transcripts from Claude)

luxicon-mcp serves a local folder of Luxicon exports to MCP clients (Claude Desktop, Claude Code) over stdio — retrieval only; the reasoning is the client's job. Export sessions from the app (per-session JSON or a person's Full History JSON) into ~/Luxicon (or pass --library <dir>; subfolder names label sessions that lack a person).

swift build -c release
claude mcp add luxicon -- "$PWD/.build/release/luxicon-mcp"

Tools: list_people, list_sessions, get_transcript, get_summary, search_transcripts, talk_time_trends. The library is re-scanned on every call, so newly pushed or AirDropped exports appear immediately.

Mac sync (push from the phone)

Instead of AirDropping exports, install the listener and pair the phone once. The easy way is the notarized installer from the Releases page — download, double-click, one admin prompt, done. From a checkout, scripts/install-listener.sh does the same via a LaunchAgent.

cat ~/Luxicon/.sync-token       # pairing token, created on first listen
# iPhone: My Voice → Mac sync → enter the token

Transcripts you push (or every new one, with auto-push) land in ~/Luxicon as JSON, ready for the MCP server. Connections are TLS-PSK on your local network; see docs/sync.md for pairing details and troubleshooting.

Tests

swift test

Consent

Recording conversations requires consent — in many jurisdictions, from all parties. Luxicon shows a reminder in the recording UI, but complying with your local law and your organization's policy is on you. Be the kind of manager who asks first.

Privacy posture

• Audio, transcripts, and voice fingerprints are stored in the app's Documents container on-device. They are included in your normal iPhone backup (encrypted by Apple; end-to-end if you use Advanced Data Protection) — so a restored phone keeps your library.
• Out of the box, the only network traffic is the model downloads — the transcription models from Hugging Face, and on iOS 26 the system speech model from Apple (no user data attached).
• On iOS 26 and later, transcription can use Apple's built-in speech model — a system component that Apple's OS downloads and runs on-device, the same way keyboard dictation works; audio still never leaves the phone.
• Opt-in features create additional traffic, all under your control:
  • Mac sync — when you pair a Mac, transcripts and summaries you push (or all new ones, if you enable auto-push) travel over your local network to that Mac, encrypted with a key derived from the pairing token. Nothing goes to the internet. See docs/sync.md.
  • Vocabulary / people URL sync — when you point the app at a vocabulary or people-roster file URL, it fetches them (https only, no cross-host redirects) when opened.
• Export is explicit: you choose what leaves the device, and when.

Contributing

Bug reports and pull requests are welcome — see CONTRIBUTING.md for dev setup, testing expectations, and the handful of things that will bite you. We follow the Contributor Covenant; report security issues privately per SECURITY.md.

License

MIT — see LICENSE. Every dependency is permissive (Apache-2.0 / MIT family); the full inventory with license and notice texts is in THIRD-PARTY-NOTICES.md, mirrored in the app under My Voice → Open-Source Acknowledgements. Model weights carry their own licenses: Pyannote segmentation (MIT), Parakeet TDT transcription (CC-BY-4.0), WeSpeaker speaker embeddings (CC-BY-4.0), and the Parakeet EOU live-caption model (NVIDIA Open Model License).