Piper TTS Online: Try It Free in Your Browser (and Check This Before You Use It Commercially)

Search for Piper TTS and nearly every result hands you a terminal command. Install Python. Download a voice. Pipe text into a binary. It works, but most people searching that term just want to hear what Piper sounds like. They don’t want a weekend project.

You can skip all of it. Piper now runs inside a normal web browser, with no install and no account, on the machine you’re reading this on.

There’s a catch most guides skip, and it isn’t technical. It’s legal. The engine is free software, but the voices aren’t all free to use the same way, and one of the most popular English voices is restricted to non-commercial use. More on that below, because it decides whether your YouTube voice-over is fine or a liability.

Try Piper right now

Open the voice generator, choose HD voice, type a sentence, and press Generate.

The first run downloads the engine and one voice, roughly 75 MB. After that, your browser keeps the files and later runs start much faster. Your text never leaves your device. The speech is created locally, and you can save it with the Download WAV button.

Open the free voice generator

What Piper actually is

Piper is a neural text-to-speech engine built to run on small, cheap hardware. It came out of Rhasspy, an offline voice-assistant project, and was tuned to run on a Raspberry Pi. That design goal explains why it’s quick and why each voice file is small enough to download in a browser.

The project changed hands recently. The original rhasspy/piper repository was archived on October 6, 2025, and development moved to OHF-Voice/piper1-gpl. That’s the Open Home Foundation, the nonprofit behind Home Assistant’s voice features. The move changed the licence too: the archived repository was MIT-licensed, and the current one ships under GPL-3.0.

For someone typing text into a web page, the GPL change barely matters. The GPL governs the software itself, not the audio you make with it (as a general rule, not legal advice). It matters a lot if you plan to build Piper into a closed-source app.

One more detail worth knowing: the new repository is openly asking for maintainers. Piper isn’t dying. It’s a community project that depends on volunteers, and you should plan around that.

Online vs installed: which one you need

If you’re generating a few minutes of audio a week, the browser wins on effort. If you’re narrating a whole book or wiring speech into an app, install it.

Quality levels, and why “medium” is usually the right pick

Every Piper voice ships in up to four tiers: x_low, low, medium and high. The difference is sample rate and model size. Low runs at 16,000 Hz with a smaller model, medium at 22,050 Hz with a smaller model, and high at 22,050 Hz with a larger model.

In practice, medium is the sweet spot. It has the same sample rate as high and a smaller download, so it suits browsers. Sizes vary by voice: one x_low voice is 27.8 MB, while the Arabic kareem voice is 63.2 MB. On mobile data, that gap matters.

Our tool picks medium automatically when a medium voice exists for your language. You can override it under Speed, pitch and voice.

The licence trap: read the MODEL_CARD first

The engine and the voices are licensed separately, and each voice inherits the terms of the dataset it was trained on. Piper’s own documentation says the voices are meant for personal use and research, and tells you to check each voice’s MODEL_CARD before using it commercially.

The lessac voice shows why. It’s one of the best-known English Piper voices, and its MODEL_CARD points to the CSTR Blizzard 2013 dataset licence, which restricts use to non-commercial research, education and demonstration. So if you put a lessac voice-over on a monetised channel or a client ad, you’re on shaky ground.

It doesn’t stop with lessac. Many other voices were fine-tuned from it. One mirror that carries only openly licensed Piper voices left out the lessac voices and every voice fine-tuned from lessac, which covers most of the English ones. Those derived voices can show a permissive licence for their own recordings while sitting on a restricted base model. The status is murky.

Some voices are clean. The ljspeech voice’s MODEL_CARD lists its dataset as public domain. The picture is also improving: a Piper maintainer has said they trained a new base model from LibriTTS-R (CC-BY) and will slowly retrain many voices so they can be used commercially.

Checking a voice takes about a minute:

1. Note the voice ID shown in our tool, for example en_US-ljspeech-medium.

2. Find that folder in the rhasspy/piper-voices repository on Hugging Face and open its MODEL_CARD.

3. Read the dataset licence line and the training line. If the voice says it was “fine-tuned from lessac,” treat it as non-commercial unless you get legal advice.

That last step is the one people skip.

Which languages work in the browser

The full Piper voice repository covers 35 languages. The browser version is narrower, and the reason is mundane. Web tools load voices from a repackaged mirror, and that mirror is a snapshot from July 5, 2024. Newer voices aren’t in it.

Hindi is the clearest gap. Piper has Hindi voices, including pratham and priyamvada, but they’re missing from the web mirror. If you pick Hindi in HD mode, our tool tells you and offers your device’s own Hindi voice instead. Most Android phones, iPhones and Windows PCs include one.

On our site, HD voice covers English, Arabic, Portuguese, Spanish and Turkish. Rather than trusting any ranking (ours included), play the same sentence through two or three voices and pick by ear.

Piper vs Kokoro: speed or polish

Kokoro is the other open model people compare with Piper. It has 82 million parameters, Apache-licensed weights, and can run fully in the browser. The Apache licence is a real advantage for commercial work, since you aren’t stuck checking a different dataset licence for every voice.

The trade-off is weight and scope. The browser build offers 28 English voices (US and UK), and the model download runs from about 90 MB to over 300 MB depending on precision. Piper loads one smaller voice at a time and speaks far more languages.

Our rule of thumb: use Kokoro (the HD English button) for English voice-overs on a computer, and Piper for other languages or when you’re on a phone.

When it doesn’t work

It’s slow on my phone. HD runs a neural network on your own device. Older phones struggle, so switch to Device voice, which plays instantly and handles long text.

The download stalls. Check your free storage first. Private or incognito windows often block the storage Piper needs, so use a normal window.

Arabic sounds off. Written Arabic often leaves out short vowels, so any engine has to guess. Adding diacritics to the words that come out wrong usually fixes them.

I want an MP3. The tool saves WAV, which every video and audio editor accepts. Convert it afterwards if you need a smaller file. More fixes are on our help page.

FAQ