EPUB to Conversational Audio: AI-Powered Podcast Generator

2 minutes read
Audio Processing • eBook Conversion • Text-to-Speech
TTS • Podcast • MLX • Books

This script extracts text from an EPUB file, formats it, and converts it into an engaging audio podcast. It uses AI-based transcript generation, rewriting, and text-to-speech synthesis to create a conversational and captivating audio output.

Here is the source code

git clone https://github.com/namuan/llm-playground.git

Pre-Requisites:

  • MLX Qwen2.5 Models:
    • These will be downloaded when the script runs for the first time.
FIRST_PASS_LLM = "mlx-community/Qwen2.5-14B-Instruct-4bit"
SECOND_PASS_LLM = "mlx-community/Qwen2.5-7B-Instruct-4bit"

Although you can configure a different MLX model if you already have it downloaded.

It uses multiple passes to generate a final transcript based on the architecture described here

  • Kokoro Onnx Models:

    wget https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files/voices.json
wget https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files/kokoro-v0_19.onnx

Please keep note of the downloaded directory and update the # CONFIG section of the script with the correct paths. See KOKORO_MODEL_VOICES_PATH and KOKORO_MODEL_PATH.

You can also try changing the Speakers by using one of the voices listed here: https://github.com/thewh1teagle/kokoro-onnx?tab=readme-ov-file#voices

Here is the complete config section from the script.

### Config

OUTPUT_DIR = Path.cwd() / "target" / "epub_pods"
OUTPUT_DIR.mkdir(exist_ok=True)

KOKORO_MODEL_VOICES_PATH = Path.home() / "models/onnx/kokoro/voices.json"
KOKORO_MODEL_PATH = Path.home() / "models/onnx/kokoro/kokoro-v0_19.onnx"
kokoro = None  # Lazy init

KOKORO_SPEAKER_1 = "bm_george"
KOKORO_SPEAKER_2 = "af_sarah"

SECOND_PASS_FILE_NAME = "second_pass.pkl"
INITIAL_PASS_FILE_NAME = "initial_pass.pkl"
FORMATTED_CHAPTER_FILE_NAME = "formatted_chapter.txt"

FIRST_PASS_LLM = "mlx-community/Qwen2.5-14B-Instruct-4bit"
SECOND_PASS_LLM = "mlx-community/Qwen2.5-7B-Instruct-4bit"

### CONFIG (END)

Once everything is set up, you can run the script by providing the path to the ePub file. Although I do recommend starting with just generated the formatted text to see if everything looks good.

./epub-to-audio.py --text-only path/to/book.epub

Then next step, you can generate a subset of chapters to check the voices and output This will only process the first 2 chapters of the book. Keep an eye on the output it currently skips chapters with < 3 lines.

./epub-to-audio.py _path/to/book.epub -v --chapters 2

All output is generated under target/epub_pods/<book-name> folder.