LLM Driven Podcast Generator

Generate a two-host educational podcast from lecture materials (PDF, PPTX, or TXT). The pipeline runs three phases: research & enrichment, scriptwriting, and audio production with OpenAI TTS. You can swap LLM providers (OpenAI or Groq), resume from intermediate artifacts and customize voices/timing in config.py.

Features

• Multi-format ingest (PDF, PPTX, TXT) with slide/page context
• Gap analysis + optional web search (DDGS) to enrich the source material
• Persona-driven scriptwriter that outputs human-readable text and JSON for TTS
• Audio producer with OpenAI TTS, noise reduction, compression, and caching
• Resumable phases: start from existing enriched content or script JSON
• Tunable presets (draft/standard/premium) and detailed knobs in config.py

Requirements

• Python 3.10+
• FFmpeg installed and on PATH (required by pydub for audio)
• Accounts/keys: OpenAI (LLM+TTS), optional Groq (LLM), optional ElevenLabs (not used by default)

Setup

python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt

Environment

Create .env with your keys:

OPENAI_API_KEY=your-openai-key
GROQ_API_KEY=your-groq-key           # optional unless using --llm-provider groq
OUTPUT_DIR=./data/outputs
QUALITY_PRESET=premium               # draft | standard | premium

Usage

From the project root after activating the venv:

# Full pipeline (OpenAI by default)
python main.py path/to/lecture.pdf --title "Quantum Basics" --notes "Focus on applications"

# Use Groq for the alternative LLM
python main.py lecture.pdf --llm-provider groq --groq-key %GROQ_API_KEY%

# Script only (skip audio)
python main.py lecture.pptx --script-only

# Resume from existing enriched content
python main.py --from-enriched data/outputs/enriched_content/enriched_20241206.txt --title "My Podcast"

# Generate audio from an existing script JSON
python main.py --from-script data/outputs/scripts/script_20241206.json

# Custom output directory
python main.py lecture.pdf --output-dir ./my_outputs

OpenAI key is always required for TTS; for script-only runs you can omit it if you pass --script-only.

Outputs

• data/outputs/enriched_content/ enriched notes (enriched_*.txt)
• data/outputs/scripts/ script text + JSON for TTS (script_*.txt/json)
• data/outputs/audio/ final mixed podcast MP3 (podcast_*.mp3)
• data/outputs/audio_cache/ cached TTS clips (if enabled)
• data/outputs/results_*.json run summary (paths, title, provider)

Configuration

config.py controls LLM/TTS models, voice selections, pacing, overlap/interruptions, caching, and quality presets. Adjust there instead of editing the agents directly.

Notes

• .gitignore already excludes .env, data/, caches, and bytecode.
• Make sure FFmpeg is installed if audio generation fails.
• License: Apache-2.0 (LICENSE).