A laptop displays an audio-to-AI workflow, with a voice recorder and notebook on a cozy desk.
A staged, mostly local voice-note pipeline: recorder, Python, whisper.cpp, Gemma 3 and Claude, with a clear line between what stays on your machine and what doesn't.

How-To Geek writer Patrick Campanale describes a personal workflow built around an offline handheld recorder, the iFLYTEK P1 Pro. He wanted a way to capture thoughts without picking up a phone and being pulled into notifications. His account is a personal build report, not a lab test. The accuracy, reliability and focus benefits he describes are his own experience and haven't been independently verified.

The design is the useful part. It splits the job into stages, each of which can be swapped out. It also shows where "local AI" stops being local.

A laptop displays an audio-to-AI workflow, with a voice recorder and notebook on a cozy desk. The pipeline as described​

Campanale says a background Python app, built with help from Codex and later Claude, watches for the recorder's volume ID to mount over USB. The stages he describes:

  1. Ingest. When the recorder mounts, the app copies recordings into a folder organized by year and month. It never deletes anything from the recorder, so the device keeps a backup. It also skips files it has already seen by comparing hashes.
  2. Transcription. whisper.cpp runs the small.en model locally. He calls the output word-for-word and rambly, filler words included.
  3. First-pass extraction. A local Gemma 3 12B model summarizes the transcript. It also proposes candidate Todoist tasks, decisions, ideas and open questions.
  4. Final processing. Claude runs as part of a morning brief. It reads both the raw transcript and Gemma's suggestions, because the smaller model sometimes misses details or misreads them. Claude then creates Todoist tasks for real commitments. Loose ideas go to an Obsidian vault. Workshop settings go into a laser-settings database.
  5. Receipts. Each created task or note carries a receipt tied to the audio hash. That prevents duplicates and lets him trace an item back to its source recording.

His morning brief also pulls in Asana assignments, overdue Todoist tasks and emails. That is part of his wider personal setup, not part of the recorder pipeline itself.

What the local pieces actually are​

The model figures he quotes check out against the whisper.cpp project's own documentation:

  • The models table lists small.en at 466 MiB on disk, matching his "around 466MB" figure.
  • Models are multilingual unless the model name includes .en, so small.en is English-only. That's fine for English dictation, but it won't help if you mix languages.
  • The project README lists a memory figure of roughly 852 MB for the small model. That is the developer's general figure, not a measurement from his machine or his recordings.
  • whisper.cpp also publishes smaller quantized variants. small.en-q5_1 is listed at 181 MiB, which could suit lower-spec machines. Campanale doesn't say he uses them.

Google's Gemma 3 model card describes the family as text-and-image input, text output models suited to tasks like summarization. That fits the transcript-in, structured-text-out role he gives it. The model card also warns that models may generate incorrect or outdated factual statements, and that open-ended or complex tasks can be hard for them. That warning supports his choice to have a second stage check the raw transcript.

Where "local" stops​

Campanale says the transcription and Gemma summary stages are offline, so audio and first-pass notes stay on his computer. That's a real privacy gain over vendor apps. He also says he no longer pairs the recorder with his phone.

The current workflow isn't fully offline, though. Claude reads the raw transcript and Gemma's suggestions to do the final sorting. Content derived from your voice therefore leaves the machine at that step. Todoist tasks are also, by nature, stored in an online service. The privacy claim holds for the early stages only.

His stated next step is to replace Claude with a local model and Hermes Agent from Nous Research, running through LM Studio on a machine with 128GB of RAM. He also mentions a Ryzen AI Max+ "Strix Halo" system with 128GB. He writes the planned model as "Qwen3.8-27B". I couldn't verify that exact name, and he says he still needs to test it. Treat the whole thing as a plan, not a finished or benchmarked deployment.

Even if that migration works, connected services such as email, calendar and Todoist are a separate question from where the model runs. Check how any agent framework reaches those services, and whether its connectors are local or hosted, before assuming nothing leaves your network.

Ideas worth borrowing​

  • Keep the stages separate. Detect, copy, transcribe, extract, then act. Each stage can be swapped, such as a different Whisper size or a different LLM, without rewriting the rest.
  • Hash everything. A stable hash does two jobs here: it skips duplicate imports, and it serves as a provenance key for downstream receipts.
  • Don't trust a summary alone. Letting the final stage see the raw transcript as well as the first model's suggestions helps when a small model drops a detail. He doesn't give error rates, so this is a sensible design, not a proven fix.
  • Never delete from the source device. Leaving originals on the recorder is cheap insurance against a bad parse.

Cautions before you copy it​

  • Check the write controls. The report says Claude creates Todoist tasks and notes. It doesn't say whether a human approves them first. Before giving any agent write access to your task lists or databases, check what review or permission controls exist.
  • Expect transcription errors. No accuracy numbers are given for his recordings, so test with your own accents, background noise and vocabulary.
  • The code isn't included here. Campanale says he published his code on GitHub, but the repository wasn't identified. Details such as the hash algorithm, error handling and record format aren't established. Review any code carefully before running it, since it watches your USB devices and writes to your task systems.

Hardware context​

iFLYTEK's store page lists the P1 Pro at $99.00 USD against a struck-through $159.00. The page's content, as retrieved here, didn't include a storage or battery spec, and a price can change at any time. The 64GB storage and 20-hour recording figures are from the article's spec box and are manufacturer-level claims. They aren't independent test results.

Bottom line​

This is a practical example of a split architecture: cheap local models handle the privacy-sensitive bulk work, and a stronger model handles the judgment calls. Campanale is candid that the cloud step is still there. Anyone copying the idea should be just as clear about it. If you want zero cloud exposure, the final stage and the task and notes services all need the same scrutiny as the transcription step.

 

References

  1. I built my own voice note pipeline to sort my ideas and tasks using local AI How-To Geek 2026-11-18T12:01:16+00:00
  2. README.md github.com
  3. whisper.cpp/models at master · ggml-org/whisper.cpp · GitHub github.com