How do you run text to speech on your own hardware?

Updated November 3, 2026.

You run text to speech on your own hardware by installing a local speech synthesis engine, downloading a voice, and converting text to audio on that machine alone. No cloud account is needed for the synthesis itself, the text never leaves the machine, and the audio can be played, saved, or used by another local program without an internet connection.

Why local speech synthesis

Local text to speech fits uses where the text should not travel, or where a connection cannot be relied on: reading notes aloud, giving a home assistant a voice, adding spoken feedback to an accessibility tool, or narrating a draft without uploading it. A cloud speech service can sound more natural in some cases, and it can offer more voices, but it requires sending the text to a server, it usually requires an account and payment, and it stops working when the connection stops. A local engine trades some voice variety for privacy, offline use, and no per-character cost, and that trade is worth it for many home and accessibility uses.

Local engines also run on modest hardware. Some are designed for CPU-only machines, including small single-board computers, and they can speak in real time, meaning the audio starts while the text is still being converted, without an awkward pause. That matters for a voice assistant, where a delay of several seconds feels broken, and for accessibility tools, where the user is waiting for the next sentence. A large language model is not needed for speech synthesis; a dedicated speech model, trained only to turn text into audio, is smaller, faster, and easier to run.

Piper: a local engine to start with

Piper is a fast, local neural text to speech engine that was created within an offline voice assistant project and is now maintained under a new repository, with a licence change that should be checked before using it in a closed product. It converts text to spoken audio using a neural model, and it runs fast enough on CPU-only hardware to speak in real time. The original repository is archived, and active development has moved to a new repository under the Open Home Foundation, so new users should start from the actively maintained repository, not the archived one, and should check its licence and release notes for the version they install.

Piper is installed with a package manager, a voice is downloaded for the language and accent wanted, and text is piped into the engine, which writes an audio file. Voices are separate downloads, and the quality, speed, and accent depend on the voice chosen, not only on the engine. Try a few voices on the same paragraph, and listen for mispronounced names, numbers, and abbreviations, because a voice that sounds good on ordinary prose may struggle with technical terms. Save the voice name and version with any audio you produce, so a later listener can reproduce it or understand why it sounds different from a newer voice.

Putting local speech into a larger setup

Local speech synthesis pairs naturally with local speech recognition, so a home assistant can both hear and speak without a cloud service. In that setup, a microphone captures audio, a local recognition model turns it into text, a local language model or a simple rule decides what to say, and a local speech engine speaks the reply. Each step runs on the same machine, and no step needs to send audio or text to a server. That chain is more work to set up than a cloud assistant, and it will not match a cloud assistant on every task, but it keeps the whole conversation at home, and it keeps working when the internet does not.

For accessibility, local speech can be added to a desktop or mobile application without a network dependency, so a user who cannot see the screen can still hear the content. Test the speech with the actual users and the actual content, because a voice that is pleasant for a short demo may become tiring over a long document, and a fast voice may be harder to understand for some listeners. Offer a speed control and a voice choice if the application allows it, and keep the text that was spoken alongside the audio, so a listener who missed a word can read it instead. Local speech is a tool for independence, not a replacement for the text itself.

Frequently asked questions

Does local text to speech need the internet?

No, not for synthesis itself, once the engine and the voice are downloaded. The internet is needed only for installation, voice downloads, and updates, not for speaking text.

Can Piper run on a small computer without a graphics card?

Yes. Piper was designed for CPU-only hardware, including small single-board computers, and it can speak in real time on such hardware. Larger speech models may need more memory, but Piper itself does not require a graphics card.

Which Piper repository should new users use?

The actively maintained repository under the Open Home Foundation, not the archived original repository. Check its licence and release notes for the version you install, because the licence changed when development moved.

Sources

Sources checked November 3, 2026. Figures that depend on your exact model, runtime or machine are estimates; check the source or test on your own hardware.

Adults only (18+). Private uncensored AI · Terms · Privacy · All guides