Podcast Creator

A tool that converts PDF content into a 2-host podcast using LLM and TTS APIs.

Overview

Podcast Creator is a comprehensive solution for transforming PDF documents and Arxiv papers into engaging, two-host podcasts. It leverages advanced LLM (Large Language Model) and TTS (Text-to-Speech) technologies to generate natural-sounding conversations between virtual hosts. The system is designed for easy deployment and integration, with a robust HTTP API interface.

Features

• PDF content extraction from uploaded files and Arxiv URLs
• LLM-powered podcast generation in a 2-host format with natural conversation flow
• OpenAI-compatible LLM and TTS integration
• WAV audio stitching for seamless podcast creation
• Dockerized deployment for easy setup and scaling
• HTTP API interface for programmatic access
• Configuration via environment variables

Planned Features

• Agentic LLM tool-use for information retrieval (e.g. date, web search, etc) for more relevant and up to date information for the podcast
• Updated UI/Frontend with more configurations exposed to control podcast generation
• Memory integration to support retrieval from "previous episodes", developed host personalities, likes/dislikes, etc

Available Endpoints

1. Create Podcast

POST /podcasts

Initiates podcast generation from PDF files or Arxiv URLs.

Request Parameters:

• files (optional if arxiv_urls exists): List of PDF files to process
• arxiv_urls (optional if files exists): List of Arxiv URLs to process

Response:

{
  "job_id": "unique-job-identifier",
  "status": "processing",
  "created_at": "timestamp"
}

2. Get Podcast Status

GET /podcasts/status/{job_id}

Retrieves the status of a podcast generation job.

Path Parameters:

• job_id: Unique identifier for the job

Response:

{
  "job_id": "unique-job-identifier",
  "status": "processing/completed/failed",
  "progress": 0-100,
  "result_file": "filename.mp3" (if completed)
}

3. Download Podcast

GET /podcasts/download/{filename}

Downloads the generated podcast file.

Path Parameters:

• filename: Filename of the podcast to download

Environment Variables

The following environment variables can be configured:

API Settings

• ALLOWED_ORIGINS: Comma-separated list of allowed origins for CORS (default: *)

Podcast Settings

• HOST_A_NAME: Name for speaker A
• HOST_B_NAME: Name for speaker B

LLM Settings

• LLM_API_HOST: URL for the LLM service (default: "http://192.168.1.16:8000")
• LLM_MODEL: LLM model to use (default: "Mistral-Small-3.2-FP8")
• LLM_HOST_TEMPERATURE: Temperature setting for LLM (default: 0.6)
• LLM_TIMEOUT: Timeout for LLM requests in seconds (default: 600)

TTS Settings

• TTS_API_HOST: URL for the TTS service (default: "http://192.168.1.16:8000")
• TTS_API_PATH: API path (default: "/v1/audio/speech")
• TTS_MODEL: TTS model to use (default: "Kyutai-TTS-Server")
• TTS_TIMEOUT: Timeout for TTS requests in seconds (default: 60)
• TTS_WAKEUP_ENDPOINT: (Optional) API ENDPOINT that will call a GET with 60 second timeout to "wake up the server"

System Settings

• AUDIO_STORAGE_PATH: Directory to store generated audio files (default: "./audio_storage")
• MAX_FILE_SIZE: Maximum allowed file size in bytes (default: 10,485,760 bytes / 10MB)

Deployment

The application can be deployed using Docker. A docker-compose.yml file is provided for easy setup.

Development

For local development, install the required dependencies:

pip install -r requirements.txt

Start the application:

uvicorn app.main:app --host 0.0.0.0 --port 8000

License

This project is licensed under the MIT License.