voice-capture

name	voice-capture
description	Transcribe Telegram voice messages, classify content, and save to Second Brain in Notion. Triggers on voice notes from Akash or Abhigna.

Voice Capture — Think Out Loud, Lyra Captures It

Step 1: Transcribe with mlx-whisper

When a voice message arrives, OpenClaw saves it as a file (usually /tmp/voice_XXXX.ogg or .oga).

# Convert to wav first (mlx-whisper works best with wav)
/opt/homebrew/bin/ffmpeg -i /tmp/VOICE_FILE.ogg /tmp/voice_transcript.wav -y -loglevel quiet

# Transcribe (tiny model is fast; use small for better accuracy on Hinglish)
/Users/akashkedia/Library/Python/3.9/bin/mlx_whisper \
  --model mlx-community/whisper-small-mlx \
  --output-format txt \
  --output-dir /tmp \
  /tmp/voice_transcript.wav

# Read result
cat /tmp/voice_transcript.txt

If the file path isn't known, check recent temp files:

ls -t /tmp/*.ogg /tmp/*.oga /tmp/*.m4a /tmp/*.wav 2>/dev/null | head -3

Hinglish note: mlx-whisper handles Hindi-English code-switching well with --language hi if needed.

Step 2: Classify

Classify transcribed text into ONE of:

• Insight — a realization or lesson ("I just realized...")
• Decision — something decided ("I've decided to...", "Going with option B")
• Idea — a new concept or feature ("What if we...", "I had an idea for...")
• Question — an open question ("I keep wondering why...", "Should I...")
• Pattern — a recurring observation ("Every time I...", "I've noticed that...")

Extract:

• Title: 5-10 word summary
• Tags from: job-hunt, relocation, content, n26, sme-lending, ai, personal, abhigna

Step 3: Save to Second Brain

NOTION_KEY=$(cat ~/.config/notion/api_key)
curl -s -X POST "https://api.notion.com/v1/pages" \
  -H "Authorization: Bearer $NOTION_KEY" \
  -H "Notion-Version: 2025-09-03" \
  -H "Content-Type: application/json" \
  -d '{
    "parent": {"database_id": "e4027aaf-d2ff-49e1-babf-7487725e2ef4"},
    "properties": {
      "Name": {"title": [{"text": {"content": "TITLE_HERE"}}]},
      "Type": {"select": {"name": "TYPE_HERE"}},
      "Source": {"select": {"name": "Voice"}},
      "Date": {"date": {"start": "DATE_HERE"}},
      "Tags": {"multi_select": [{"name": "TAG1"}]},
      "Notes": {"rich_text": [{"text": {"content": "FULL_TRANSCRIPTION_HERE"}}]}
    }
  }'

Step 4: Confirm

Reply: "Captured. [Type]: [Title] → Second Brain ✓"

Example: "Captured. Idea: Auto-draft Substack from X posts via Lyra → Second Brain ✓"

Notes

• Always save full verbatim transcription in Notes, not just the summary
• Abhigna voice notes: tag "abhigna", still save to Second Brain (Akash's database)
• If transcription fails (file not found, ffmpeg error), ask Akash to resend and explain why
• Clean up temp files after: rm -f /tmp/voice_transcript.wav /tmp/voice_transcript.txt