estebanstifli/LocalText2Voice

A complete local production workflow for clean narration, structured learning content, and podcast-ready audio

LocalText2Voice – What It Is

LocalText2Voice is a free, open‑source desktop application that lets you turn long pieces of text (books, lessons, articles, notes, etc.) into spoken‑word audio files. It runs on Windows and Linux and can work entirely offline using locally‑installed text‑to‑speech (TTS) models, or it can optionally call cloud TTS services such as OpenAI, ElevenLabs, Google Gemini, or Azure Speech.

Why It Matters

Privacy‑first – When you use a local engine, your source text never leaves your computer.
No subscription needed – All core features work with free, downloadable models; cloud APIs are optional.
Built for long‑form content – The workflow handles chapters, headings, and large documents, splitting them into safe chunks for reliable generation.
Quality‑control loop – An optional Faster‑Whisper step can transcribe the generated audio, compare it to the original text, and flag mismatches for review or automatic retry.
Podcast‑ready output – After narration you can add background music, apply fades, ducking, and loudness normalization, then export a single polished file.

Core Workflow (11 steps)

Import text – paste or load *.txt, *.md, .docx files.
Optional normalization – apply language‑specific dictionaries to make the text pronunciation‑friendly.
Smart chunking – the app breaks the document into paragraph‑aware segments.
LTV markup (optional) – embed commands like {{voice "Emma"}}, {{pause 900ms}}, {{speed 0.9}} directly in the source.
TTS generation – choose a local engine (Piper, Kokoro, Chatterbox, Qwen3‑TTS, OmniVoice, optional F5‑TTS Russian) or a cloud API; the engine runs per‑segment.
Clean narration – the generated WAV files are concatenated into a single audio file (M4B, MP3, M4A, Opus, FLAC, OGG).
Optional Whisper review – Faster‑Whisper transcribes each segment, computes similarity/WER, and marks segments for approval or retry.
Manual/automatic fixes – edit, regenerate, or trim problematic segments.
Subtitle export – create .srt or karaoke‑style .ass files using Whisper timestamps.
Audio mix – add background music, set volumes, apply fades, ducking, and loudness normalization without re‑running TTS.
Export – final audio (plus optional subtitles, project metadata, cover art) is written to disk.

Main Features

Feature	What It Does
Modular TTS engines	Supports many local models (Piper, Kokoro, Chatterbox, Qwen3‑TTS, OmniVoice, F5‑TTS Russian) and cloud APIs. Models are downloaded on‑demand and isolated in their own runtimes.
Voice library & cloning	Browse a catalog, download Piper voices, or clone new voices from a short reference recording (Chatterbox, OmniVoice, F5‑TTS).
LTV Markup	Inline commands to switch voice, language, pause length, speed, volume, or pass model‑specific instructions.
Text normalization	Language‑specific dictionaries (Arabic, Chinese, English, …, Russian) that rewrite numbers, dates, currencies, etc., before TTS.
Whisper review	Faster‑Whisper can transcribe generated audio, compare it to the source, compute similarity scores, and automatically retry low‑quality segments.
Subtitle generation	Produces `.srt` and word‑level `.ass` files, correctly offset for mixes.
Audio mix page	Add background music, control volumes in dB, set intro/outro offsets, enable ducking, and normalize loudness for podcast distribution.
Bulk audiobooks	Queue multiple books, assign individual covers/music, and generate them sequentially with pause/resume/retry support.
Project persistence	All project data (metadata, segment list, voice choices, Whisper scores, timestamps) stored in SQLite and a portable manifest for easy re‑opening or future extensions.

Supported Platforms & Requirements

Windows 10/11 (64‑bit) – official installer (LocalText2Voice-Setup.exe).
Linux (64‑bit) – run from source; needs Python 3.10+, a virtual‑env, and FFmpeg in PATH.
macOS – not officially supported yet.
Hardware – works on modest CPUs (4‑core, 8 GB RAM) for lightweight engines; GPU (NVIDIA CUDA) optional but speeds up larger models like Chatterbox, Qwen3‑TTS, OmniVoice.

Licensing & Cost

Application – MIT License, fully free.
Local models – free to run, but each model follows its own upstream license (e.g., CC‑BY‑NC for OmniVoice, non‑commercial CC‑BY‑NC 4.0 for F5‑TTS Russian).
Cloud APIs – optional; usage is billed by the provider.

Getting Started (Windows Quick‑Start)

Download the latest installer from the Releases page.
Run the installer and pick a folder for the app and a separate data folder for AI assets.
Choose a setup profile – CPU light (Piper) or Powerful GPU (OmniVoice + Faster‑Whisper).
Open Settings → TTS Engines, install the desired engine, and download a voice.
Paste or import your text, hit Generate Audio, and (if enabled) review Whisper results.
Use the Audio Mix tab to add music and export the final file.

Who Might Use It?

Authors & educators who want to self‑publish audiobooks or lecture recordings without paying for commercial TTS services.
Podcasters looking for a privacy‑preserving way to generate narration and mix it with music.
Developers & hobbyists interested in experimenting with open‑source TTS models and building custom voice pipelines.

Bottom line: LocalText2Voice bundles a complete, offline‑first pipeline for turning long‑form text into high‑quality spoken audio, with optional cloud fallbacks, quality‑control via Whisper, and built‑in podcast‑mixing tools—all under an MIT‑licensed, community‑driven codebase.