Building a Local-First AI Recording App to Escape Cloud-Locked Recorders

Problems with Existing Recording Cards

The author investigated mainstream magnetic recording cards and found four major issues: hardware prices starting at 499 CNY for entry-level models and exceeding 1000 CNY for pro versions; closed ecosystems that lock recordings into vendor apps; mandatory cloud transcription that uploads private audio to third-party servers; and ongoing subscription fees for transcription minutes plus inability to use external AI APIs like Kimi or DeepSeek.

Hardware Choice: 199 CNY Open-Protocol Card

A 199 CNY magnetic recording card was selected for its openness. It uses standard BLE (Bluetooth Low Energy) with public protocol documentation, measures 4.32 mm thick, includes 32 GB storage, ~20 hours battery life, and combines a silicon MEMS microphone with a bone-conduction microphone for reliable standalone recording.

Because the card exposes its BLE protocol, the author could write a native iOS app that connects directly via Bluetooth, syncs recordings to the phone, and processes everything locally without any vendor cloud.

Core Decision: Pure On-Device Offline Processing

Modern iPhones include Apple GPU and ANE (Neural Engine), providing strong on-device inference capability. Since users already paid for this silicon when buying the phone, the app avoids paying again for cloud compute.

Local offline transcription + speaker diarization : Recordings are transcribed on-device with high accuracy, automatically separating speakers (Speaker Diarization) and supporting on-device voiceprint profiles to recognize familiar voices.

Google Gemma 4 running locally via Apple MLX : To generate high-quality meeting notes without cloud, the app integrates the Gemma 4 model using Apple's MLX framework.

Works in airplane mode : Both verbatim transcripts and structured summaries (key conclusions, TODOs) are produced entirely offline; audio and text never leave the device.

Because processing is local, there are no transcription time limits — whether recording a 10-minute idea or a 1000-hour interview, no quota is consumed.

Multiple Input Sources: A Universal Voice Inbox

The app accepts audio from four local sources, making it useful even without the hardware card:

Magnetic card sync : Record during meetings, then open the app to Bluetooth-sync and auto-transcribe.

Apple Watch wrist recording : Independent recording on Watch, auto-transfer to iPhone when nearby.

iPhone direct recording : Quick capture via app or system Shortcuts.

Local audio/video import : Import any audio file; for video, the app automatically extracts the audio track (e.g., DASH AAC from Bilibili) without downloading the full video.

Podcast & Long-Video "One-Click Dehydration"

When listening to podcasts on Xiaoyuzhou, Himalaya, Apple Podcasts, or Bilibili, users can share the link to Lynsight. The app parses the audio stream (for Bilibili it directly extracts the independent DASH AAC track, saving bandwidth), downloads it, then runs local transcription and AI summarization, turning a 30-minute episode into a structured, timestamped note in minutes.

User-Controlled AI & Data Export

Unlike products that lock both AI choice and data export, Lynsight lets users decide both:

Bring your own LLM (Kimi / DeepSeek / Qwen / any OpenAI-compatible API) : The built-in Gemma 4 handles daily notes privately. For deeper long-form extraction, users can enter their own API keys. Even when using a remote model, only the transcript text is sent after explicit confirmation; original audio and voiceprints stay local.

Open delivery to Obsidian & Apple Reminders : Transcripts, structured summaries, and conclusions can be one-way written to a specified Obsidian vault (preserving later manual edits), action items sync to Apple Reminders, and standard Markdown, TXT, or timestamped SRT subtitles can be exported anytime.

Anti-Subscription Pricing: 7-Day Free Trial, One-Time Purchase

No account required : Download and use immediately.

7-day full-feature free trial : No auto-charge after trial ends.

One-time buyout : Single payment (roughly the price of a simple meal) unlocks the full version forever — no monthly/yearly fees or minute top-ups.

Data always yours : Even if the trial expires without purchase, previously recorded audio and generated text remain accessible, playable, editable, and exportable.

Development Retrospective (2 Weeks from Zero to App Store)

The author built the entire native iOS app — covering BLE hardware communication, Apple Watch coordination, on-device speech recognition, and local LLM inference — in just over two weeks. Key challenges included:

Transferring large recording files over BLE: solving disconnections, checksum verification, and throughput.

Initial local Qwen model failed for summarization; switched to Gemma 4 after on-device blind testing.

Running on-device LLMs caused SIGKILL due to iOS memory limits; worked around Apple's default memory constraints.

Implementing a zero-server, account-free trial and buy-once IAP that passed strict App Store review and filing.

A follow-up article will detail the full development-to-release process.