Podcastfy Generator 🎙️

Generate AI podcast-style audio conversations from any content. Creates engaging two-host dialogues similar to Google NotebookLM's Audio Overview feature.

Capabilities

• URLs → Fetch article content, generate podcast discussion
• YouTube → Extract transcript, create audio summary
• PDFs → Parse document, synthesize key points as dialogue
• Text/Topics → Generate podcast from plain text or topic prompts
• Multi-lingual → English, German, French, Spanish (auto-detect or specify)
• Custom Identity → Name the podcast, name the hosts, pick their voices

Quick Examples

"Create a podcast about this article: https://example.com/tech-news"
"Turn this YouTube video into a podcast: https://youtube.com/watch?v=..."
"Generate a German podcast discussing quantum computing"
"Make a podcast called 'Deep Dive' with hosts Alex and Sam about this PDF"

Usage

Basic Generation

# From URL
<skill>/scripts/generate.py --url "https://example.com/article"

# From YouTube
<skill>/scripts/generate.py --url "https://youtube.com/watch?v=abc123"

# From text
<skill>/scripts/generate.py --text "Your content here..."

# From PDF
<skill>/scripts/generate.py --pdf "/path/to/document.pdf"

# Multiple sources
<skill>/scripts/generate.py --url "https://url1.com" --url "https://url2.com"

Podcast Identity

# Name the podcast
<skill>/scripts/generate.py --url "https://..." --podcast-name "Deep Dive"

# Name the hosts (they'll use each other's names in conversation)
<skill>/scripts/generate.py --url "https://..." --host-name Alex --cohost-name Sam

# No podcast name (hosts introduce topic naturally, no show branding)
<skill>/scripts/generate.py --url "https://..." --podcast-name ""

# Full customization
<skill>/scripts/generate.py --url "https://..." \
  --podcast-name "Tech Talk" --podcast-tagline "Breaking down the future" \
  --host-name Alex --cohost-name Kiki

Language Options

# Auto-detect (default)
<skill>/scripts/generate.py --url "https://example.de/artikel"

# Explicit language
<skill>/scripts/generate.py --url "https://example.com" --lang de

Supported: en (English), de (German), fr (French), es (Spanish)

TTS Provider & Voice Options

Default: OpenAI TTS (tts-1-hd with onyx + nova voices)

Optional: ElevenLabs for higher quality, more natural voices:

# Use ElevenLabs with defaults (Daniel + Alice)
<skill>/scripts/generate.py --url "https://..." --elevenlabs

# Custom voices per host
<skill>/scripts/generate.py --url "https://..." --elevenlabs \
  --host-voice Daniel --cohost-voice Alice

# OpenAI custom voices
<skill>/scripts/generate.py --url "https://..." \
  --host-voice echo --cohost-voice shimmer

# Use local sherpa-onnx TTS (free, offline, unlimited)
<skill>/scripts/generate.py --url "https://..." --sherpa

OpenAI voices: alloy, echo, fable, onyx, nova, shimmer

ElevenLabs voices (premade): Roger, Sarah, Laura, Charlie, George, Callum, River, Liam, Alice, Matilda, Will, Jessica, Eric, Bella, Chris, Brian, Daniel, Lily, Adam, Bill

Sherpa-onnx (local): Uses Piper VITS models. Voice paths configured in config/conversation.yaml under text_to_speech.sherpa. Requires sherpa-onnx-offline-tts binary (set SHERPA_ONNX_TTS_BIN or install to ~/.openclaw/tools/sherpa-onnx-tts/). Performance note: CPU-based synthesis, typically ~2-10x realtime, requires ~2GB+ RAM, and quality is good but generally below ElevenLabs.

Browse ElevenLabs voices: https://elevenlabs.io/voice-library

All CLI Options

Option	Description	Example
--url	URL to process (repeatable)	--url https://...
--text	Plain text content	--text "AI is..."
--pdf	Path to PDF file	--pdf report.pdf
--lang	Output language	--lang de
--podcast-name	Podcast name (empty = none)	--podcast-name "Deep Dive"
--podcast-tagline	Podcast tagline	--podcast-tagline "..."
--host-name	Host name (Person1)	--host-name Alex
--cohost-name	Co-host name (Person2)	--cohost-name Kiki
--elevenlabs	Use ElevenLabs TTS	--elevenlabs
--sherpa	Use local sherpa-onnx TTS (free)	--sherpa
--host-voice	Voice for host	--host-voice Daniel
--cohost-voice	Voice for co-host	--cohost-voice Alice
--output, -o	Output file path	-o podcast.ogg

Output

The script outputs an OGG audio file path. Use the OpenClaw message tool to send it:

# Agent workflow
audio_path = exec("<skill>/scripts/generate.py --url 'https://...'")
message(action="send", media=audio_path, target=user_chat)

Configuration

Default podcast style is configured in <skill>/config/conversation.yaml. CLI flags override config values.

Key config options:

• podcast_name — Show name (empty = content-driven intro)
• roles_person1 / roles_person2 — Host role descriptions
• text_to_speech.{provider}.default_voices — Default voice per provider
• language_voices.{provider}.{Language} — Per-language voice overrides (applied when no --host-voice/--cohost-voice is set)
• conversation_style — Style keywords (engaging, concise, etc.)
• creativity — 0-1 scale (higher = more creative dialogue)

Environment Variables

Variable	Required	Purpose
OPENAI_API_KEY	Yes	TTS audio generation (default)
GEMINI_API_KEY	Yes	Transcript/dialogue generation
ELEVENLABS_API_KEY	No	ElevenLabs TTS (required for --elevenlabs)
SHERPA_ONNX_TTS_BIN	No	Path to sherpa-onnx-offline-tts binary (for --sherpa)

Get your ElevenLabs API key at: https://elevenlabs.io/app/settings/api-keys

Installation

First-time setup (run once):

<skill>/scripts/install.sh

Requirements

• ffmpeg — Audio format conversion
• uv — Python environment management
• Python 3.11+ — Runtime

Troubleshooting

"ffmpeg not found"

Install ffmpeg: brew install ffmpeg (macOS) or apt install ffmpeg (Linux)

"API key not set"

Ensure OPENAI_API_KEY and GEMINI_API_KEY are in your environment or secrets.conf

Hosts say "Quick Brief" or reference a show name

Set podcast_name: "" in config/conversation.yaml or use --podcast-name ""

Generation takes too long

Podcastfy processes content through LLM + TTS. Expect 30-90 seconds for short podcasts.

Audio quality issues

Try ElevenLabs (--elevenlabs) for more natural voices. OpenAI tts-1-hd is decent but synthetic.