Developer-readable App data: none. No account, no sign-up. HarkNote processes recordings, transcripts, speaker labels, optional Self voice matching, summaries, and Ask on your phone. It does not send legal acceptance records or recording content to CTK Advisors LLC. This page covers the iPhone and Android apps and their Apple Watch and Wear OS companions; where they differ, it says so.
iPhone. Recording, transcription (Whisper via WhisperKit or NVIDIA Parakeet via FluidAudio — your choice, both Core ML / Apple Neural Engine), speaker identification, and optional Self voice matching run entirely on your iPhone. Summaries and Ask use Apple's built-in on-device model on iPhones with Apple Intelligence enabled, and a built-in deterministic text heuristic otherwise.
Android. Recording and transcription (NVIDIA Parakeet v3 by default, or OpenAI Whisper, both run through sherpa-onnx) execute entirely on your phone. Speaker identification, if you turn it on, runs pyannote segmentation and a WeSpeaker voice-embedding model through sherpa-onnx on the device. Summaries and Ask use Gemma 4 E2B or Qwen3.5-2B through Google's LiteRT-LM runtime where your phone can run them, and the same deterministic heuristic everywhere else.
Either way, no audio or transcript text is sent off-device for processing.
Watches. On Apple Watch and Wear OS, HarkNote records completely offline — no phone nearby, no internet. Finished sessions transfer to your paired phone when it is back in range and are transcribed there. On Apple Watch that is Apple's private device-to-device connection. On Wear OS it is Google's Wear Data Layer, which uses Bluetooth or, when no direct link is available, a Google relay that Google documents as end-to-end encrypted between your two devices. Neither path is a developer-readable upload or cloud AI processing.
No user content uploads to us. HarkNote may download on-device model assets so transcription, speaker detection, and summaries can run locally — from Hugging Face on iPhone, and from the sherpa-onnx GitHub releases and Hugging Face on Android, verified against pinned hashes. These requests do not include your audio, transcripts, notes, questions, summaries, calendar content, voice profiles, or speaker embeddings.
iPhone only: App database data, including optional voice profiles and speaker embeddings, may sync via Apple's private CloudKit service when iCloud is available for HarkNote. Audio files use your iCloud Drive container only if you select iCloud storage. We never receive or can read your iCloud data. Android has no sync service; everything stays in the app's private storage on that phone.
When you accept Privacy & Terms, HarkNote stores the document versions, timestamp, app/build, and document hashes only in local preferences so it can require renewed acceptance after an update. The App does not send that record to us.
When you share a transcript, summary, or SRT captions, the app you pick receives it and handles it under its own terms. Deleting a session in HarkNote does not recall copies you have exported.
If you separately submit feedback through harknote.app, the website receives the message and optional contact information you choose to provide. Do not include recording or meeting content in website feedback.
No analytics. No telemetry. No crash uploads. No advertising IDs. No trackers.
Microphone
Required only to record audio you start. Never accessed otherwise.
Calendar
Optional, read-only. Links a recording to a meeting. HarkNote never creates, edits, or deletes events. On Android, if you separately enable Wear OS meeting reminders, the next meeting's title and start time are sent to your paired watch; that option is off by default.
Notifications (Android)
Optional. Shows the recording in progress so you can see and stop it from the shade. On iPhone the same job is done by a Live Activity, which needs no permission.
All of these can be declined or revoked in iOS or Android Settings at any time.
Original audio is kept until you delete it by default. Automatic deletion is an explicit choice and starts only after Ready; incomplete work remains protected, and Keep Original Audio always wins for that session.
Resumable processing uses local job manifests and temporary transcript-section checkpoints in the App sandbox. On iPhone they are excluded from backup and never enter CloudKit or iCloud Drive; on Android app backup is disabled for both the phone and Wear apps. Content-bearing checkpoints are removed after transcript and notes commit safely.
On iPhone, a content-free local diagnostic journal may remain for up to 30 days. It contains operational facts and sanitized codes, never session content or stable identifiers. HarkNote uploads nothing automatically; you preview a Private Support Report before explicitly copying it. Android keeps local operational diagnostics for recovery only and integrates no crash-upload service.
Delete any session in-app — HarkNote removes its processing artifacts and audio before deleting transcript and metadata. Uninstalling removes local data from that device. On iPhone, recordings stored in your selected iCloud Drive container are standard .caf files accessible through Files; local-only recordings remain inside the App sandbox. On Android, recordings are standard WAV files in the app's private storage. We keep no server-side copies of your recording content.
Recording-consent laws vary by jurisdiction and you are solely responsible for complying with them — obtaining consent and giving notice where required. See the Privacy Policy for the full notice.
iPhone speech-to-text: WhisperKit (Argmax, MIT) or NVIDIA Parakeet via FluidAudio (framework Apache-2.0; Parakeet models CC-BY-4.0), plus their open-source dependencies under MIT / Apache-2.0 licenses.
Android speech-to-text: sherpa-onnx (Apache-2.0) running NVIDIA Parakeet TDT 0.6B v3 (models CC-BY-4.0, converted by sherpa-onnx) or OpenAI Whisper (MIT). Speaker identification: pyannote segmentation 3.0 (MIT) and the WeSpeaker VoxCeleb ResNet34-LM voice-embedding model (CC-BY-4.0), both through sherpa-onnx. Summaries and Ask: Gemma 4 E2B and Qwen3.5-2B (Apache-2.0), run through Google's LiteRT-LM runtime.
Typefaces, used under the SIL Open Font License 1.1 on both platforms: DSEG7 Classic, Silkscreen, JetBrains Mono, IBM Plex Mono, VT323, Anton, Poiret One, Rye, Pacifico, Major Mono Display, Press Start 2P, Audiowide, UnifrakturMaguntia. Special Elite is used under the Apache License 2.0.
Typeface license files ship inside the app bundle. The library and model licenses are published at their canonical URLs: MIT — opensource.org/license/mit · Apache-2.0 — apache.org/licenses/LICENSE-2.0 · CC-BY-4.0 — creativecommons.org/licenses/by/4.0 · SIL Open Font License 1.1 — openfontlicense.org