LocalText2Voice

Open-source, offline AI text-to-speech for long-form content.
Convert documents into MP3 narration, podcasts, audiobooks, lessons, and course audio on Windows.

Download for Windows · Listen to an MP3 demo · AndromedaNova.com

LocalText2Voice is a lightweight Python and PySide6 desktop application that runs neural text-to-speech locally with Piper TTS. It is designed for long documents, works without a GPU, and uses FFmpeg to produce clean narration or a complete podcast mix with music, fades, loudness normalization, and ducking.

Resumen en español: aplicación de escritorio open source que convierte textos largos en voz y podcasts MP3 mediante inteligencia artificial local. Funciona sin conexión, sin GPU y con voces Piper descargables desde la propia interfaz.

Screenshots

Generate long-form audio without freezing the interface

Browse, preview, download, and manage multilingual Piper voices

Why LocalText2Voice?

• Private by default: text and speech synthesis stay on the computer.
• Built for long content: safe text chunking, chapter detection, paragraph pauses, and sequential output files.
• No GPU required: Piper provides fast neural TTS inference on normal CPUs.
• Podcast-ready: keep clean narration and optionally create a second mix with intro, background music, outro, fades, normalization, and ducking.
• Portable Windows app: extract a folder, run the executable, and download a voice from the built-in manager.
• Open architecture: new local or cloud TTS engines can implement one stable interface without rewriting the application.

Typical uses include course narration, study materials, accessibility, podcasts, audiobooks, training content, articles, documentation, and voice-over drafts.

Features

Offline AI speech synthesis

• Local neural TTS with Piper and ONNX voice models.
• Built-in voice catalog with language/quality filters.
• Remote voice sample playback before downloading a model.
• Background downloads with cancellation, size validation, SHA-256 checks, and atomic installation.
• Voice speed control and automatic discovery of installed models.
• No cloud account, API key, GPU, or global Piper installation required.

Long-form text processing

• Paste text or import .txt, .md, and .docx files.
• Normalize whitespace and unsupported characters.
• Preserve paragraph boundaries and split long text into TTS-safe blocks.
• Detect Markdown headings, chapters, lessons, modules, and short uppercase headings.
• Export one MP3 or one MP3 per chapter.
• Natural randomized paragraph pauses with adaptive timing after long paragraphs and periodic reading breaks.

Podcast production

• Export clean narration as podcast1.mp3, podcast2.mp3, and so on without overwriting previous work.
• Create a separate podcast mix while retaining the clean narration.
• Optional intro, looped background music, and outro from MP3 or WAV files.
• Music volume, fade in/out, and silence between sections.
• Basic FFmpeg sidechain ducking while the voice is speaking.
• Optional loudness normalization to -16 LUFS.
• Configurable MP3 bitrate and ID3 title, artist, and album metadata.

Desktop experience

• Responsive PySide6 interface with worker-thread generation.
• Live block progress, elapsed time, estimated remaining time, and visible log.
• Safe cancellation of the active Piper process and temporary-file cleanup.
• Open-output-folder action after generation.
• English, Spanish, French, German, Italian, Portuguese, Simplified Chinese, Japanese, Arabic, and Hindi interface translations.
• Right-to-left layout support for Arabic.

Quick Start on Windows

1. Open the latest release.
2. Download LocalText2Voice-v0.3.0-windows-x64.zip.
3. Extract the ZIP to a folder where you have write permission.
4. Run LocalText2Voice.exe.
5. Open Manage voices, preview a voice, and download it.
6. Paste or import your text, choose the voice, and select Generate Audio.

The portable release includes the application, Piper runtime, and FFmpeg. Voice models are downloaded separately because their licenses and dataset conditions can differ.

Audio example

Listen to the English MP3 demo generated locally with Piper TTS

The demo was generated by LocalText2Voice with the en_GB-alan-medium Piper voice. It contains no cloud-generated speech.

How It Works

flowchart LR
    A["TXT, Markdown, DOCX, or pasted text"] --> B["Text normalization and heading detection"]
    B --> C["Safe chunks and natural pause plan"]
    C --> D["Piper neural TTS / ONNX inference"]
    D --> E["Temporary WAV blocks"]
    E --> F["FFmpeg assembly and MP3 encoding"]
    F --> G["Clean narration"]
    F --> H["Optional podcast mix"]

Loading

AI Engineering Highlights

LocalText2Voice is an applied AI engineering project rather than a model training project. It integrates pretrained neural speech models into a reliable end-user workflow:

• Local model inference: orchestrates Piper ONNX voices through an isolated subprocess adapter with executable and model validation.
• Model lifecycle management: discovers a remote model catalog, previews samples, validates downloads, and installs voices atomically.
• NLP-oriented preprocessing: applies language-agnostic text cleanup, sentence-aware chunking, heading heuristics, and prosody-oriented pause planning for long-form synthesis.
• Production concurrency: keeps the GUI responsive with Qt worker threads, progress signals, ETA calculation, cancellation, and child-process cleanup.
• Audio DSP pipeline: composes narration and music with FFmpeg filters for looping, fading, sidechain compression, metadata, and loudness normalization.
• Extensible providers: isolates TTS behind BaseTTSEngine and reserves a provider interface for future LLM-assisted course generation.

This demonstrates practical experience with AI model integration, ONNX-based inference, desktop product engineering, multimedia processing, asynchronous workflows, internationalization, packaging, and automated testing.

Technology Stack

Technology	Role
Python 3.10+	Application, processing pipeline, configuration, and tests
PySide6 / Qt 6	Native desktop interface, signals, threads, and i18n-ready UI
Piper TTS	Fast local neural text-to-speech engine
ONNX voice models	Portable pretrained speech models
FFmpeg	WAV assembly, MP3 encoding, mixing, ducking, fades, and loudnorm
PyInstaller	Portable Windows folder distribution
python-docx	Optional Microsoft Word document import

Architecture

LocalText2Voice/
|-- main.py
|-- app/
|   |-- core/       # Text processing, settings, project and audio pipeline
|   |-- tts/        # Engine interface, Piper adapter and voice management
|   |-- ui/         # PySide6 windows and reusable widgets
|   |-- workers/    # Background generation and download workers
|   |-- utils/      # Paths, logging and FFmpeg helpers
|   `-- llm/        # Future provider interface; no cloud integration yet
|-- locales/        # Auto-discovered JSON translations
|-- engines/piper/  # Portable Piper runtime
|-- voices/         # External ONNX voice models
|-- ffmpeg/         # Portable FFmpeg executable
|-- music/          # Optional intro, background and outro library
|-- tests/
`-- output/

The audio pipeline depends on the abstract BaseTTSEngine contract. Piper is the first implementation; Kokoro, XTTS, ElevenLabs, OpenAI TTS, Azure TTS, or other providers can be added later without coupling them to the UI.

Run from Source

Python 3.10 or newer is recommended:

git clone https://github.com/estebanstifli/LocalText2Voice.git
cd LocalText2Voice
py -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
python main.py

The repository excludes large third-party binaries and voice models. Place the Windows Piper runtime in engines/piper/ and FFmpeg in ffmpeg/, or point the application settings at compatible local installations. Voices can be installed from Manage voices.

Expected Piper files:

engines/piper/
|-- piper.exe
|-- required DLL files
`-- espeak-ng-data/

Expected voice pair:

voices/es_ES/davefx/medium/
|-- es_ES-davefx-medium.onnx
`-- es_ES-davefx-medium.onnx.json

FFmpeg is resolved from ffmpeg/ffmpeg.exe first and then from the Windows PATH. MP3 export requires the libmp3lame encoder.

Configuration

The application creates config.json beside the executable or source project. If the file is missing, safe defaults are created automatically. config.example.json documents the available values, including:

• output folder, voice, language, speed, split mode, and export mode;
• Piper and FFmpeg paths;
• chunk size and block/chapter pauses;
• randomized and adaptive paragraph pause rules;
• MP3 bitrate, metadata, and clean-audio normalization;
• intro, background, outro, fades, gaps, ducking, and podcast normalization.

Build the Portable App

build_windows.bat

The build uses PyInstaller --onedir mode and creates:

dist/LocalText2Voice/
|-- LocalText2Voice.exe
|-- engines/
|-- voices/
|-- ffmpeg/
|-- music/
|-- output/
|-- licenses/
|-- config.example.json
|-- LICENSE
`-- THIRD_PARTY_NOTICES.md

Piper, voices, and FFmpeg remain outside the main executable so they can be updated independently. Do not redistribute a voice until you have reviewed its model card and dataset license. Third-party license texts are copied into the portable folder under licenses/.

Tests

python -m unittest discover -s tests -v

The test suite covers text normalization, safe splitting, chapter detection, settings, output naming, natural pauses, localization, and voice catalog behavior. End-to-end synthesis additionally requires real Piper, voice, and FFmpeg files.

Roadmap

• Offline Piper TTS generation for long-form text
• Downloadable voice catalog with remote previews
• Clean narration and music-backed podcast exports
• Natural pauses, progress, ETA, cancellation, and multilingual UI
• Short product video and animated workflow demo
• Visual chapter editor before synthesis
• More local engines such as Kokoro and XTTS
• Optional cloud providers such as OpenAI TTS, Azure, and ElevenLabs
• PDF and richer document import
• Optional LLM-assisted lesson and course generation
• Signed Windows installer, automatic updates, and release automation

Contributing

Issues, feature proposals, translations, and pull requests are welcome. Read CONTRIBUTING.md before submitting a change. Please never add large voice models, copyrighted music, API keys, or generated build folders to the repository.

License

LocalText2Voice source code is released under the MIT License. Piper, FFmpeg, PySide6/Qt, voice models, and other dependencies keep their own licenses. See THIRD_PARTY_NOTICES.md before redistributing a portable build.

Author

Created by Esteban at AndromedaNova.com.

If LocalText2Voice helps your work, starring the repository makes the project easier to discover.