Introducing Loudkit: speech on your own hardware

I want speech to be something you can put inside your own software. A reader, a study tool or an accessibility feature should be able to speak without sending each passage to a hosted speech service. With Loudkit, you can inspect the code, keep the model locally and decide how speech fits into the rest of your product.

LoudReader brings that approach to reading books, articles and documents. Releasing the engine opens it to ideas beyond a reading app. You supply the interface and the workflow; Loudkit supplies the voice.

Start with text and save spoken audio, or stream longer passages as they are generated. The SDKs let you bring that into the language your project already uses. Python has the reference implementation; Swift, Go, Rust and TypeScript have their own guides and runtime requirements.

There are two models: loudr-1 and loudr-1-turbo. They share the voice-profile format and public API. Begin with the default model, then compare Turbo using the text you actually want people to hear. The model guide explains the choice without treating one speed measurement as a promise for every device.

After the initial model download, speech generation stays on your machine. You do not need a speech-service account or a per-character allowance. You do need room for the model and the dependencies for your chosen runtime. Those requirements vary by SDK and backend, so check the supported platforms and limits before choosing a deployment target.

Local speech is one part of an application. If your app fetches a web page or sends a prompt to an online model, those actions still use the network. Loudkit gives you control over where the speech itself happens.

Both model releases include 28 voices across 10 languages. Listen before choosing: English has been evaluated by ear; the other nine languages have automated checks but still need native-speaker review. Feedback on pronunciation and naturalness is useful, especially on the material you plan to read.

You can also make a portable voice profile from roughly five to ten seconds of clean speech. Use your own recording or obtain the speaker's permission. The voice-cloning guide covers the additional models and input requirements. A saved profile works with either synthesis model.

The Python quickstart is a short route to a first audio file. In a Python environment that meets the installation guide's requirements, run:

pip install "loudkit[torch,audio,hub]"
loudkit speak --voice joe "Hello from Loudkit." --play -o hello.wav

The first run fetches the model files. The command saves a WAV and asks your system player to play it. From there, try a passage from your own project and another voice. That is a better first test than judging a speech engine from a feature list.

For a ready-made reading experience, LoudReader provides natural offline voices on iPhone, iPad and compatible Apple Silicon Macs. Loudkit is for building your own experience. And Loudkit for agents is a separate companion preview for talking to an agent you already use. That preview currently needs an Apple Silicon Mac; its setup page explains which integrations are tested. These are three ways into the same idea: speech that can run on your own hardware.