AI Podcast Generator

Transform PDFs into engaging AI-powered podcasts using Azure services.

Overview

This application takes PDF documents and converts them into natural-sounding podcast episodes with structured narratives. It uses:

• Azure Blob Storage: Store PDFs and generated podcast audio (with Managed Identity support)
• Azure AI Search: Index and retrieve PDF content with vector embeddings
• Azure OpenAI: Generate podcast scripts (GPT-4o) and synthesize audio (gpt-4o-audio-preview)
• Azure Key Vault: Secure secrets management (optional)
• Azure Functions: Serverless deployment option

Features

• 📄 PDF text extraction with multiple fallback methods
• 🔍 Semantic search with vector embeddings (text-embedding-3-large)
• ✍️ Structured script generation with sections (Intro, Segments, Outro)
• 🎙️ Section-based audio synthesis for proper narrative flow
• 🔐 Managed Identity authentication for enterprise security
• 🎵 Proper MP3 metadata with accurate duration tracking
• ☁️ Cloud-native Azure integration with DefaultAzureCredential
• 🎨 Multiple podcast styles: conversational, educational, summary, storytelling

Project Structure

KJPodcastApp/
├── src/
│   ├── services/
│   │   ├── blob_storage.py       # Azure Blob Storage operations
│   │   ├── pdf_processor.py      # PDF text extraction
│   │   ├── search_service.py     # Azure AI Search integration
│   │   ├── script_generator.py   # Podcast script generation
│   │   └── audio_synthesizer.py  # Audio generation
│   ├── config.py                  # Configuration management
│   ├── orchestrator.py            # Pipeline orchestration
│   └── cli.py                     # Command-line interface
├── azure_function/                # Azure Function deployment
├── pdf-index.json                 # Azure AI Search index schema
├── requirements.txt
├── setup.py                       # Setup verification script
├── .env.example                   # Environment configuration template
├── README.md
└── GETTING_STARTED.md

Prerequisites

• Python 3.9+
• Azure CLI
  • Windows: winget install --exact --id Microsoft.AzureCLI or download from Microsoft Learn
• FFmpeg (required for audio processing)
  • Windows: winget install Gyan.FFmpeg or download from gyan.dev/ffmpeg
  • macOS: brew install ffmpeg
  • Linux: sudo apt-get install ffmpeg
• Azure subscription with services provisioned:
  • Azure OpenAI (GPT-4o, text-embedding-3-large, gpt-4o-audio-preview)
  • Azure Blob Storage
  • Azure AI Search

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Configure Azure authentication:

   Option A: Managed Identity (Recommended for Azure environments)

   cp .env.example .env
   # Edit .env and set the following:
   AZURE_KEY_VAULT_URL=https://your-key-vault.vault.azure.net/
   AZURE_STORAGE_ACCOUNT_NAME=your-storage-account
   AZURE_OPENAI_API_KEY=your-api-key
   AZURE_OPENAI_GPT4O_DEPLOYMENT=your-gpt4o-deployment
   AZURE_OPENAI_EMBEDDING_DEPLOYMENT=your-embedding-deployment
   AZURE_OPENAI_AUDIO_DEPLOYMENT=your-audio-deployment
   AZURE_SEARCH_ENDPOINT=https://your-search.search.windows.net
   AZURE_SEARCH_API_KEY=your-search-api-key
   AZURE_SEARCH_INDEX_NAME=pdf-index
   # Optional: AZURE_MANAGED_IDENTITY_CLIENT_ID=your-client-id
   # Optional: AZURE_STORAGE_CONNECTION_STRING=your-connection-string-here

   Option B: Local Development with Azure CLI

   az login
   # Then configure .env with storage account name

3. Create Azure AI Search Index:

   The application requires a search index with vector search capabilities. You can create it using:

   Option A: Automatic (Recommended) - The index is created automatically on first run

   Option B: Manual - Use the provided schema:

   # Upload the index schema to Azure AI Search
   az search index create --service-name <search-service> \
     --name pdf-index \
     --schema @pdf-index.json

   See pdf-index.json for the complete index schema with vector search configuration.

4. Verify setup:

   python setup.py

5. Generate your first podcast:

   python -m src.cli generate document.pdf --output podcast.mp3

Usage

Generate Podcast from PDF

python -m src.cli generate document.pdf --output podcast.mp3

Upload PDF to Blob Storage

python -m src.cli upload document.pdf

Process Existing PDF in Blob Storage

python -m src.cli process --blob-name document.pdf

Azure Function Deployment

The application can be deployed as an Azure Function for serverless execution:

cd azure_function
func azure functionapp publish <your-function-app-name>

Architecture

Pipeline Flow

1. PDF Ingestion: Upload PDFs to Azure Blob Storage with Managed Identity authentication
2. Text Extraction: Extract and clean text from PDFs using pdfplumber/PyPDF2
3. Vector Indexing: Create searchable index with text-embedding-3-large embeddings in Azure AI Search
4. Structured Script Generation: GPT-4o creates podcast scripts with clear sections (Introduction, Segments, Outro)
5. Section-Based Audio Synthesis: Each script section is synthesized independently using gpt-4o-audio-preview
6. Audio Combination: Sections are combined with pydub to create properly formatted MP3 with correct metadata
7. Storage: Save podcast audio to Blob Storage with content-type metadata

Key Technical Features

• Managed Identity Support: Uses DefaultAzureCredential for secure, keyless authentication
• Vector Search: Semantic search with 3072-dimensional embeddings using HNSW algorithm
• Section-Based Synthesis: Scripts are split by logical sections (Intro/Segments/Outro) for proper narrative ordering
• Base64-Encoded Keys: Search document keys use URL-safe base64 encoding to handle special characters
• Proper MP3 Metadata: pydub ensures combined audio files have accurate duration and headers

Azure AI Search Index Schema

The application uses a custom index schema (pdf-index.json) with the following fields:

Field	Type	Description
id	String	URL-safe base64-encoded document key (handles special characters)
blob_name	String	Original PDF filename in blob storage
content	String	Text chunk content (searchable)
title	String	Document title
chunk_index	Int32	Chunk sequence number for ordering
num_pages	Int32	Total pages in source PDF
embedding	Collection(Single)	3072-dimensional vector from text-embedding-3-large

Vector Search Configuration:

• Algorithm: HNSW (Hierarchical Navigable Small World)
• Metric: Cosine similarity
• Parameters: m=4, efConstruction=400, efSearch=500
• Optimized for fast approximate nearest neighbor search

Development

• Python 3.9+
• FFmpeg for audio processing
• Azure subscription with required services provisioned
• Azure CLI for local development authentication
• Local development with Azure Functions Core Tools (optional)

License

MIT