🎧 Text2Audio

Turn a book into a natural-sounding, chaptered audiobook — for free, fully on your own machine.

🔊 Hear a sample — a short clip narrated by Text2Audio.

Text2Audio is a local, open-source audiobook studio. Drop in your manuscript (.txt or Markdown), pick a narrator, and get a mastered .m4b audiobook with chapter markers — no cloud, no accounts, no per-use cost. It can even clone a voice from a short audio sample.

It runs entirely offline on your own GPU using open text-to-speech models (Kokoro-82M by default; F5-TTS for voice cloning).

Screenshots

Create	Voices & cloning	In-app player

---

🚀 Quick start (no terminal needed — Windows)

Don't want to touch a command line? On Windows you can be up and running in three steps:

1. Download the project as a ZIP and unzip it somewhere (for example, your Desktop).
2. Double-click Install Text2Audio.bat. It sets everything up for you — Python, the app and its dependencies, ffmpeg, and espeak-ng. The first run takes a while; just let it finish.
3. Double-click Start Text2Audio (the installer also adds a shortcut to your Desktop). Your browser opens the Studio — start making audiobooks. Keep the little black window open while you use the app; closing it stops Text2Audio.

💡 The first time you generate audio, the voice model (a few hundred MB) downloads automatically. After that, everything runs fully offline.

⚠️ Windows may show a blue "Windows protected your PC" SmartScreen box on the .bat files — click More info → Run anyway (see below). The scripts are plain text you can open and read first. A full walkthrough, voice-cloning setup, and troubleshooting live in INSTALL.md.

Install with pip (any OS, for technical users)

pip install git+https://github.com/mooja77/Text2Audio
text2audio

Add the optional voice-cloning engine with the clone extra:

pip install "text2audio[clone] @ git+https://github.com/mooja77/Text2Audio"

You'll still need ffmpeg and espeak-ng installed on your system, and — for GPU speed — a CUDA build of PyTorch (pip pulls the CPU build by default). See Setup below for the details.

---

Features

• 📖 Book → audiobook — .txt / Markdown in, chaptered .m4b out (plays in Apple Books, Smart AudioBook Player, BookPlayer, etc. with chapter navigation).
• 🖥️ Studio web UI — drag-and-drop multi-file import, a voice gallery, live per-chapter progress, a built-in chapter player, and a library of your renders.
• 🎚️ Mastered audio — 128 kbps AAC, loudness-normalized (~−19 LUFS), high-pass filtered, with natural sentence/paragraph pacing. Re-master any book instantly (no re-synthesis).
• 🗣️ Reads text correctly — expands numbers/abbreviations and an editable pronunciation dictionary fixes tricky names.
• 🎤 Voice cloning (optional) — upload ~10–30 s of speech and narrate in that voice, via F5-TTS (runs isolated; Kokoro stays the fast default).
• 🔒 100% local & free — your text and audio never leave your machine.

Requirements

• Windows (developed/tested on Windows 11; should adapt to Linux/macOS) with an NVIDIA GPU + recent driver (CPU works but is slow).
• Python 3.10–3.12
• ffmpeg on your PATH (provides ffmpeg and ffprobe).
• espeak-ng (Windows .msi) — phonemizer backend for Kokoro.

Setup

Non-technical user on Windows? Skip this — use the Quick start above and INSTALL.md instead. The steps below are the manual developer setup.

git clone https://github.com/mooja77/Text2Audio.git
cd Text2Audio
python -m venv .venv
.\.venv\Scripts\Activate.ps1
# Install a CUDA build of PyTorch first (pick the index for your CUDA version):
pip install --index-url https://download.pytorch.org/whl/cu124 torch
pip install -r requirements.txt

If espeak-ng isn't auto-detected, point to it:

setx PHONEMIZER_ESPEAK_LIBRARY "C:\Program Files\eSpeak NG\libespeak-ng.dll"
setx PHONEMIZER_ESPEAK_PATH "C:\Program Files\eSpeak NG\espeak-ng.exe"

Linux / macOS

The project is plain Python and cross-platform. The commands are the same, except:

• Activate the venv with source .venv/bin/activate.
• Install ffmpeg and espeak-ng from your package manager:
  • Debian/Ubuntu: sudo apt install ffmpeg espeak-ng
  • macOS (Homebrew): brew install ffmpeg espeak-ng
• Where this README shows .\.venv\Scripts\python.exe, use .venv/bin/python.
• Pick the PyTorch index for your platform/accelerator from pytorch.org (e.g. ROCm on Linux, or the default CPU/MPS build on macOS — note Kokoro/F5 are much faster on CUDA).

Run

.\.venv\Scripts\python.exe server.py

Your browser opens the Studio. Create tab: drag in .md/.txt chapter files (Markdown is auto-cleaned), set title/voice/speed, and Generate — live per-chapter progress streams as it renders. Voices: audition narrators or clone one. Library: every finished audiobook with an in-app chapter player, re-master, and delete.

A classic single-screen UI is also available: python app.py.

Chapter markers

For one-file-per-chapter imports, each file's first # Heading becomes the chapter title. For a single .txt, put a marker line before each chapter:

## Chapter One
Once upon a time...

## Chapter Two
...

If there are no markers, it auto-detects Chapter N / ALL-CAPS headings, else treats the whole text as one chapter.

Voice cloning (optional)

Voice cloning uses F5-TTS, a heavier optional engine. Kokoro stays the default; cloning runs in an isolated subprocess (F5 and Kokoro can't share a process).

.\.venv\Scripts\python.exe -m pip install f5-tts
# f5-tts may pull a mismatched torchaudio — re-pin it to match your torch:
.\.venv\Scripts\python.exe -m pip install "torchaudio==2.6.0" --index-url https://download.pytorch.org/whl/cu124

Then in Voices → Clone a voice: name it, upload ~10–30 s of clean speech, and (optionally) paste a transcript of the clip for best quality. The cloned voice appears alongside the built-ins and in the Create dropdown. Cloned renders are higher quality but much slower — use ▶ Sample to audition first.

⚠️ Responsible use

Only clone voices you have the right to use — your own voice, voices you have explicit permission to clone, or public-domain recordings. Do not impersonate people or create deceptive audio. You are responsible for how you use this tool.

Tests

The suite runs without a GPU (the TTS engines are mocked):

pytest -q

GPU model tests are opt-in: $env:RUN_KOKORO=1; pytest tests/test_synth.py and $env:RUN_F5=1; pytest tests/test_clone_synth.py.

How it works

book (.txt/.md) → ingest (clean markdown) → normalize (numbers/abbr/pronunciation)
   → chunk (sentence/paragraph) → synthesize (Kokoro or F5) → assemble (ffmpeg:
   chapters + loudness master) → .m4b in your library/

Project layout: pipeline/ (text→audio), backend/ (library, jobs, render, voices, pronunciations), server.py (FastAPI + web API), web/ (vanilla-JS UI).

Contributing

See CONTRIBUTING.md. Issues and PRs welcome.

License

Text2Audio's code is MIT (see LICENSE). The TTS models it downloads have their own licenses — notably the F5-TTS cloning weights are CC-BY-NC (non-commercial). See THIRD_PARTY_LICENSES.md.

Acknowledgements

• Kokoro-82M — default narration
• F5-TTS — voice cloning
• ffmpeg, espeak-ng, FastAPI, and the wider open-source ecosystem