Piper
Fast, local neural text-to-speech engine with downloadable voices in dozens of languages, run from the command line.
Turn text into speech on your own machine with downloadable voices, a CLI, a web server and a Python API.
Piper is a neural text-to-speech engine that runs locally and embeds espeak-ng for phonemisation. It suits developers, accessibility projects and home-automation builders who want offline voices rather than a paid speech API. It is a command-line and library tool, not a studio with a timeline.
What it is
You install it with pip, download a voice by name, and it writes a WAV file or plays audio directly. The same engine is available as an HTTP server with a small test page, a Python API and a C/C++ library. The repository also contains the code to train or fine-tune your own voices.
Who it's for
It fits people wiring speech into their own software, screen-reader setups or smart-home projects, and anyone who wants narration without sending text to a cloud service. It does not suit you if you want voice cloning from a short sample, a point-and-click editor, or clear commercial rights to every voice.
Key features
Local neural synthesis
Generates speech on your own computer from a downloaded voice model, with optional --cuda GPU acceleration.
Dozens of languages
The voices page lists languages from Arabic and Catalan to Welsh, Hindi and Swahili.
CLI, HTTP and Python
Use the command line, a built-in web server with a test page and /synthesize endpoint, or the Python and C/C++ APIs.
Raw phoneme injection
Insert espeak-ng phonemes inline with [[ ]] blocks to fix a mispronounced word.
Train your own voices
Training code in src/piper/train fine-tunes from an existing checkpoint using a pipe-delimited CSV of clips and text.
Per-voice licence guidance
A commercial-use page lists voices whose whole training chain permits commercial use.
Up and running in 3 steps
Our honest verdict
What's great
- The HTTP server has a browser page that shows the voice name, language, speaker list and the phonemes with audio alignments for the last utterance.
- You can adjust length, noise and noise W scales per request, and pick a speaker on multi-speaker voices.
- Inline [[ phoneme ]] blocks let you correct a single word without retraining anything.
- The project documents exactly which voices have a commercially clean training chain instead of leaving you to guess.
Where it falls short
- There is no graphical editor or timeline: you work from a terminal, scripts or the HTTP API, and the CLI reloads the model on every call, so repeated use needs the web server.
- The project grants no licence over voice models. Each voice carries its own dataset terms, some voices are non-commercial, and you must check the training chain yourself before commercial use.
- The engine is GPL-3.0-or-later, so shipping it inside a closed-source product needs a licence review.
- The README says the Open Home Foundation is looking for maintainers, so ask whether you are comfortable depending on it long term.
Pricing & license
Piper is free to download and run under its GPL licence, and the project's own pages show no paid plan. The only cost is your own compute, plus any rights questions you resolve for the voice you choose.
Open remote AI-training roles
870 open remote AI-training roles, updated November 14, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I use Piper voices commercially?
The engine's licence allows it under GPL terms, but the project grants no licence over the voices. Its commercial-use page lists voices whose entire training chain uses commercially permitted data, and says to check the MODEL_CARD of the copy you downloaded.
Does Piper need a GPU?
No. It runs on CPU by default, and the CLI has a --cuda option if you install onnxruntime-gpu.
Can it clone my voice?
Not from a short sample. The repository includes training code that fine-tunes a voice from a CSV of audio clips and transcripts, ideally starting from an existing checkpoint.
