Memoria: Keeping Family Voices Alive with Local Whisper + Gemma

Description and use

Memoria is a local oral‑history archivist that transcribes and summarizes family voice memos using Whisper and Gemma, keeping all data on the local machine.

Problem

Family voice memos contain sensitive personal data that should not be sent to cloud services for transcription and summarization.

Outcome

Memoria achieves fully local, privacy-preserving transcription and structured summarization of family voice memos, producing JSON and Markdown archives within ~30 s on a local CPU.

Workflow

• Upload or record a family voice memo
• Run Whisper small (CPU int8) to generate transcript with timestamps
• Feed transcript to Gemma 3 4B (llama.cpp) to produce structured fields (title, summary, people, places, time references, themes, quotes, follow-up questions)
• Apply deterministic quote verification to ensure quotes exist in transcript
• Export results as JSON and Markdown
• Optionally view via the FastAPI web interface

Evidence

Memoria runs both speech recognition and language understanding locally. The recording and transcript never leave the machine for inference.

Source

The bundled 58‑second story is synthetic and fictional, narrated with TTS specifically so that no private family recording is exposed.

Source

I ran the project rather than stopping at a screenshot or a plausible demo response: Unit tests 6/6 passed, demo audio duration 57.79 seconds, full CLI run 33.36 seconds.

Source