Podcast Audio Processing

Skill name: podcast-audio-processing

Process podcast audio from NotebookLM: convert to mp3, transcribe with Whisper, create chapters, and embed metadata.

---

When to Use This Skill

Use after receiving audio from NotebookLM (Phase 9 in workflow):

1. User has downloaded audio from NotebookLM (typically .m4a or .wav)
2. Episode needs: conversion, transcription, chapters, metadata embedding

---

How to Invoke

Use the Task tool with subagent_type="general-purpose" and prompt:

"Process the podcast audio file for this episode using the podcast-audio-processing skill.

Episode path: podcast/episodes/YYYY-MM-DD-topic-slug
Audio filename: [filename user provided, e.g., 'Original_Audio.m4a']
Episode slug: YYYY-MM-DD-topic-slug

Follow the podcast-audio-processing skill to:
1. Convert to mp3 if needed (m4a → mp3)
2. Get file metadata (size in bytes, duration)
3. Transcribe with local Whisper (base model)
4. Analyze transcript and create 10-15 chapter markers
5. Embed chapters into mp3

CRITICAL: Report back the file metadata when complete:
- Duration: MM:SS format
- File size: bytes"

---

Workflow

Step 1: Convert Audio Format (if needed)

Check if the audio file is .m4a or .wav format. If so, convert to mp3:

cd ~/src/research/podcast/episodes/EPISODE_PATH

# Convert to mp3 (128kbps for optimal size/quality)
ffmpeg -i "AUDIO_FILENAME.m4a" -codec:a libmp3lame -b:a 128k "EPISODE_SLUG.mp3" -y

Note the metadata from ffmpeg output:

• Duration (format: HH:MM:SS or MM:SS)
• This will be needed for publishing

Step 2: Get File Metadata

Get the file size in bytes:

ls -l EPISODE_SLUG.mp3 | awk '{print $5}'

Record:

• File size: [bytes]
• Duration: [from ffmpeg output]

Step 3: Generate Transcript with Local Whisper

Run Whisper transcription locally (no API key needed):

cd ~/src/research/podcast/tools

# Basic transcription (uv run auto-manages dependencies)
uv run python transcribe_only.py ../episodes/EPISODE_PATH/EPISODE_SLUG.mp3 --model base

# OR with organized logging (recommended for production)
mkdir -p ../episodes/EPISODE_PATH/logs
uv run python transcribe_only.py ../episodes/EPISODE_PATH/EPISODE_SLUG.mp3 \
  --model base \
  --log-dir ../episodes/EPISODE_PATH/logs \
  --quiet

Whisper model options:

• tiny: Fastest (~1-2 min for 30 min audio), basic accuracy
• base: [recommended] Fast (~5-10 min), good accuracy
• small: Slower (~15-20 min), better accuracy
• medium: Slowest (~30-40 min), best accuracy

Default to base model unless user specifies otherwise.

Output:

• Creates: EPISODE_SLUG_transcript.json in the episode directory
• With --log-dir: Also creates timestamped log file in logs/ directory
• --quiet: Suppresses progress messages (useful in automated workflows)

Step 4: Analyze Transcript and Create Chapters

Read the transcript file and analyze it to identify natural topic transitions.

Chapter creation guidelines:

• Aim for 10-15 chapters for a 30-40 minute episode
• Each chapter should be 2-4 minutes long
• Chapter titles should be descriptive and capture the key topic/story
• Include subtitles or key concepts after the main title when helpful
• Analyze the full transcript to identify natural topic transitions
• Look for: topic shifts, new concepts introduced, story transitions, framework changes

Create two chapter files:

1. FFmpeg metadata format (EPISODE_SLUG_chapters.txt):

;FFMETADATA1
[CHAPTER]
TIMEBASE=1/1000
START=0
END=120000
title=Introduction: The Topic Overview

[CHAPTER]
TIMEBASE=1/1000
START=120000
END=300000
title=Historical Context: Early Development

2. Podcasting 2.0 format (EPISODE_SLUG_chapters.json):

{
  "version": "1.2.0",
  "chapters": [
    {
      "startTime": 0,
      "title": "Introduction: The Topic Overview"
    },
    {
      "startTime": 120,
      "title": "Historical Context: Early Development"
    }
  ]
}

Important format notes:

• FFmpeg format: START/END in milliseconds (TIMEBASE=1/1000)
• Podcasting 2.0 format: startTime in seconds (decimal)
• Last chapter END time should match audio duration in milliseconds

Step 5: Embed Chapters into MP3

Embed the chapter metadata into the mp3 file:

cd ~/src/research/podcast/episodes/EPISODE_PATH

# Embed chapters using FFmpeg metadata file
ffmpeg -i EPISODE_SLUG.mp3 -i EPISODE_SLUG_chapters.txt -map_metadata 1 -codec copy temp.mp3 -y

# Replace original with chaptered version
mv temp.mp3 EPISODE_SLUG.mp3

Result:

• Chapters embedded in mp3 file
• Will appear in podcast apps that support chapters (Overcast, Pocket Casts, Apple Podcasts)
• File size remains the same

---

First-Time Setup

# Fix SSL certificates (macOS Python - one-time)
/Applications/Python\ 3.12/Install\ Certificates.command

# Dependencies auto-managed by uv - no manual install needed
# Just use: uv run python transcribe_only.py ...

---

Technical Notes

Whisper Transcription

• Transcription is 100% local: No API calls, completely private, free
• Transcript files are large: 300-400KB - read in sections if needed
• Whisper timing: base model takes ~5-10 minutes for 30-minute audio

Chapter Support

• Modern podcast apps will display chapters; older apps will ignore them

---

Files Created

After completion, these files should exist in the episode directory:

File	Size	Description
EPISODE_SLUG.mp3	~30MB	Final audio with embedded chapters
EPISODE_SLUG_transcript.json	~400KB	Full transcript with timestamps
EPISODE_SLUG_chapters.txt	~2KB	FFmpeg chapter format
EPISODE_SLUG_chapters.json	~1KB	Podcasting 2.0 format

---

Error Handling

Audio Conversion Errors

If conversion fails:

• Check that audio file exists and path is correct
• Verify ffmpeg is installed: ffmpeg -version

Transcription Errors

If transcription fails:

• Always use uv run python (not bare python) - this ensures correct venv
• Verify audio file exists and path is correct
• Try smaller model (tiny) if base is too slow

Chapter Embedding Errors

If chapter embedding fails:

• Verify chapters.txt file exists and has correct format
• Check that START/END times don't exceed audio duration
• Ensure TIMEBASE is set correctly (1/1000 for milliseconds)

---

Final Report

When complete, report back to main agent:

• Audio processed: EPISODE_SLUG.mp3
• Duration: MM:SS
• File size: [bytes]
• Transcript generated: [N] words
• Chapters created: [N] chapters
• Chapters embedded in mp3
• Files ready for publishing phase