Sloth Secretary 🦥: I Made Gemma Take the Meeting Minutes So I Could Take a Nap

Description and use

Sloth Secretary automatically transcribes and summarizes meeting recordings offline with Whisper and Gemma, requiring no API keys or cloud costs.

Problem

The secretary role requires transcribing and summarizing long call recordings, which is time‑consuming and often done with cloud APIs that incur costs, data usage, and privacy concerns.

Outcome

Sloth Secretary processed a 63‑minute pod call in 22 minutes on a CPU‑only laptop, producing a draft with 8 tasks correctly attributed and a 2‑line TL;DR. The tool required no cloud access, avoided data costs, and allowed quick human review before sharing.

Workflow

• 1. Convert call recording to 16 kHz mono WAV with ffmpeg.
• 2. Transcribe with faster‑whisper (small model, CPU, int8) and VAD to skip silences.
• 3. Split long transcripts into ~10‑minute chunks with overlap for map‑reduce.
• 4. Send each chunk to Gemma via Ollama or llama‑cpp‑python to generate structured JSON notes.
• 5. Merge chunk notes, dedupe, fold TL;DR, and add a "⚠️ Please double‑check" section.
• 6. Output a Markdown file and a ready‑to‑paste WhatsApp/Telegram message.
• 7. Optionally run in watch mode to auto‑process new recordings.

Evidence

Give it an audio or video file from any call app and it writes: a 2-line TL;DR, decisions, open questions, action items grouped by person, a Markdown file plus a ready‑to‑paste WhatsApp/Telegram message

Source

All of it runs on my laptop: open Whisper weights for the listening, Gemma for the writing. There's no API key, nothing gets uploaded, and nobody sends me a bill.

Source

Gemma 3 4B caught all 8 tasks and gave each one the right owner. It also put the projector room on Wednesday (it's Thursday), dropped the Sunday deadline on Emeka's Supabase task, and missed one decision...

Source

Real numbers, measured in a CPU‑only sandbox (2 ARM vCPUs, 8 GB RAM, no GPU): ... The real pod call (63‑min file) ... 21 min 56 s

Source

The sloth earned its name the hard way, ... Gemma 4B got OOM‑killed on chunk 3 of 6. Now Whisper runs in a child process and hands its memory back before Gemma loads...

Source