[Feature Request / Contribution] Mobile Voice-to-Vault Pipeline: Local Whisper STT & "Inbox" Workflow

[gnekt]

I am a PhD student in AI and an Engineer, currently managing a high volume of meetings and research notes. I’ve been looking for a streamlined way to bridge the gap between "on-the-go" thoughts/recordings and a structured Obsidian vault without the friction of manual porting.

Your roadmap (v3.1–v3.4) covers incredible ground, but I believe there is a gap in mobile-first data ingestion. I am very interested in this project and would like to contribute a workflow that transforms the phone into a "remote control" for the vault.

What I've in my mind

I propose (and offer to develop) an integration or a side-utility that follows this logic:

• Local STT: Integration of models like Whisper (via whisper.cpp or similar) for private, local speech-to-text.

• The "Remote Control" App: A lightweight interface to record, transcribe, and instantly push the output to a specific GitHub directory (e.g., /inbox/to-process).

• The Processing Loop: Once home, these files can be picked up by the desktop environment (or an automated Claude Desktop skill) for final sorting and Knowledge Graph integration.

---

Questions for Maintainers

Do you have plans for local STT (Whisper/Faster-Whisper) within the current architecture?
Is a "Mobile Ingestion" workflow already on your radar for the v3.2 or v3.4 milestones?

[m13v]

this is a workflow I use daily, basically the same pipeline you describe. voice recording on the go, local Whisper STT, then structured ingestion into a knowledge store.

a few things from building this: whisper.cpp is the right call for mobile/local STT. the quantized models (especially distil-large-v3) give you near-cloud accuracy without needing a network connection. for the mobile capture piece, the biggest friction point is not the transcription itself but the metadata - you want timestamps, location, and ideally some context about what you were doing when you recorded it.

for the processing loop, we built something that ingests raw transcripts plus browser history and other signals into a SQLite database with embeddings for semantic search. the trick is making the vault-side processing smart enough to tag and route notes automatically instead of requiring manual sorting. even a simple keyword extraction step saves a lot of time.

[m13v]

the knowledge extraction and storage layer we built for ingesting transcripts and browser data into a searchable database: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/db.py

and the embeddings module for semantic search over ingested content: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/embeddings.py