README.md

The production pipeline for My Weird Prompts, a semi-automated AI podcast that combines human-recorded prompts with AI-generated responses, rendered as listenable audio episodes.

Note: This validated pipeline has been integrated into the main My Weird Prompts website deployment repository at danielrosehill/My-Weird-Prompts.

About the Show

My Weird Prompts is a podcast where the host shares interesting, unusual, or thought-provoking prompts and AI responds with detailed, conversational explanations. The result is a hybrid human+AI podcast format that preserves the authenticity of the original question while leveraging AI for comprehensive responses.

The Approach

This workflow addresses several limitations with existing AI podcast solutions:

Why Not Notebook LM?

Notebook LM produces great content, but the podcast style tends toward a specific Americanized, California-esque host format that doesn't suit everyone's preferences.

Why Not Pure TTS Workflows?

Previous N8N workflows (speech-to-text → LLM → text-to-speech) work but have issues:

• Quality TTS (like ElevenLabs) is expensive for regular 30-minute episodes
• Fully synthetic output lacks the human element

Production Pipeline

A hybrid approach:

1. Human prompts - Recorded audio prompts from the creator
2. AI responses - Multimodal AI generates podcast-style dialogue
3. Combined output - Final episode includes: intro jingle → disclaimer → human prompt → AI dialogue → outro jingle

Workflow

[Audio Prompt Recording]
        ↓
[Gemini: Transcribe + Generate Script]
        ↓
[Chatterbox TTS via Replicate]
        ↓
[Assemble Episode]
   - Intro jingle
   - Disclaimer
   - Human prompt audio
   - AI dialogue (~15 min)
   - Outro jingle
        ↓
[Render Normalized MP3 + Metadata]
        ↓
[Upload to Spotify]

Quick Start

# Drop audio prompts into the queue
cp your-prompt.mp3 pipeline/prompts/to-process/

# Generate episodes
./generate.sh

# Output appears in pipeline/output/episodes/<episode-name>/

Generation Process

Here's what the episode generation looks like in action:

---

1. Startup & File Detection

The generator scans for audio prompts in the queue:

---

2. Voice Sample Upload & Audio Processing

Voice samples are uploaded to Replicate while the prompt audio is processed (silence detection, trimming):

---

3. Script Generation

Gemini transcribes the prompt and generates a ~15-minute diarized podcast script:

---

4. AI Dialogue Preview

The generated script features natural conversation between hosts Corn and Herman:

---

5. TTS & Cover Art Generation

Parallel workers generate voice-cloned audio segments while Flux Schnell creates cover art variants:

---

6. Episode Assembly & Normalization

All 32 dialogue segments are concatenated, then the full episode is assembled with loudness normalization to -16 LUFS:

---

7. Generation Complete

Final episode with auto-generated description, ready for upload:

---

8. Output Files

Each episode gets its own folder with the final MP3, cover art variants, script, and metadata:

---

9. Generated Metadata

Full episode metadata including title, description, and technical details:

---

Cost

Approximately $0.40 per 15-minute episode (Replicate TTS + Gemini + cover art generation).

Repository Structure

• pipeline/generators/ - Episode generation scripts
• pipeline/prompts/ - Audio prompt queue (to-process/done)
• pipeline/output/episodes/ - Rendered episodes
• pipeline/show-elements/ - Intro/outro jingles
• config/voices/ - Voice samples for TTS cloning