TTS

Convert a markdown file into an audio narration using VCTK and Coqui-TT

A Bash script that converts Markdown files into narrated audio using Text-to-Speech (TTS) via Docker and the Coqui-TTS framework.

Features

• Converts a Markdown (.md) file to MP3 audio.
• Uses the VCTK multi-speaker TTS model with customizable voices.
• Cleans Markdown syntax (headings, links, code blocks, etc.) for more natural-sounding speech.
• Splits large files into manageable chunks for processing.
• Optional playback of the generated audio.

Prerequisites

• Docker (with GPU support recommended for faster synthesis)
• Docker Compose
• Python 3 (for model invocation and Markdown cleaning)
• ffmpeg (for audio processing)
• mp3wrap (for merging audio chunks)
• str (for string manipulation)

Installation

1. Clone the repository:

   git clone https://github.com/ngpepin/TTS.git
   cd narrate-md
   

2. Make the script executable:

   chmod +x narrate-md.sh
   

3. (Optional) Place the script in your PATH for global access:

   sudo ln -s $(pwd)/narrate-md.sh /usr/local/bin/narrate-md

Usage

   ./narrate-md.sh [-p] <input.md>

Options

| Flag | Description                          |
| ---- | ------------------------------------ |
| `-p` | Play the generated audio immediately |

Example

./narrate-md.sh -p README.md  # Converts README.md to README.mp3 and plays it

Configuration

Edit these variables in the script to customize behavior:

MODEL="tts_models/en/vctk/vits"  # TTS model (default: VCTK multi-speaker)
SPEAKER="p230"                   # Speaker ID (e.g., p230-p260 for VCTK)
MAX_LINES_PER_CHUNK=20           # Split large files into chunks of this size

Technical Details

Pipeline

1. Markdown Cleaning:

   • Strips headings, links, code blocks, and formatting

   • Adds pauses for punctuation (e.g., commas → ",,")

2. Audio Generation:

   • Uses Coqui-TTS in a Docker container

   • Converts text to WAV → MP3 with tempo adjustment

3. Chunk Handling:

   • Splits large files (>20 lines by default)

   • Merges chunks into a single MP3

Supported Models

• VCTK: Multi-speaker English (113 speakers, various accents)

• LJSpeech: High-quality single-speaker English (via Tacotron2)

License

MIT License. See LICENSE for details.

---

Warning
This is a work in progress. Error handling is minimal, and results may vary with complex Markdown.