FireVoice

Overview

The problem

The best-sounding voice tools are cloud services. You pay per character, your scripts and voice samples go to someone else's server, and nothing works offline. Open voice models can match them, but each one needs its own Python setup, the right graphics card and a lot of trial and error before it says a single line.

What it does

• Speak

  Write a line, pick a voice and a delivery style such as storyteller, trailer or whisper, and FireVoice performs it, with pauses, laughs and breaths placed in the text

• 15 engines, one app

  From Inferno, the flagship 48 kHz engine with cloning and voice design in 30 languages, to Spark, which runs instantly on a CPU, so you can switch engines per script

• 100 ready-made voices, plus cloning from a short reference clip and voices designed from a written description

• Dialogue, dubbing and a voice changer for multi-speaker scripts and existing recordings

• Dictation in any program

  Hold a global hotkey, speak, and the words appear wherever your cursor is, in a live or a polished mode, with voice commands, punctuation, a personal dictionary, replacements and snippets

• Transcribe recordings to text locally

• Guided first run

  A six-step setup reads your graphics card, RAM and free disk space and recommends the engine that fits

• A local voice API, no keys required

  An HTTP API with built-in OpenAPI docs and OpenAI- and ElevenLabs-style routes, so existing scripts, apps and AI agents can use the same voices by changing one address

• Projects, jobs and stats to organise long work, plus signed updates that install without blocking you

How it's built

• A Tauri 2 app with a Rust shell and a React 19, Vite and Tailwind CSS 4 interface
• A Python FastAPI backend that the app bundles and runs, which hosts every engine, the job queue and a streaming dictation socket
• Runs on Windows 10 and 11 and on Linux as an AppImage, .deb or .rpm package, on an NVIDIA GPU or on the CPU
• The release ships no readable source: the backend is compiled with Cython, the interface is minified and the binary is stripped
• The landing site is built with Next.js 16, Tailwind CSS 4, React Three Fiber with a particle field of about 100,000 points, GSAP and Lenis, a generative Web Audio score and real renders from the app. It scores 100 on all four Lighthouse categories on desktop

Result

Gallery

Links

• Live sitefirevoice.fabwaseem.com
• Download for Windowsfirevoice.fabwaseem.com/download/windows
• Releasesgithub.com/fabwaseem/firevoice-releases/releases