Chinese Text to Speech: Free, In-Browser, Nothing Uploaded

Mandarin is the exact inverse of Korean here — the HD engine is the one that cannot speak it. Paste Chinese text, press play, and the audio is generated on your own device — no account, no upload, no character limit. This page is the honest version: which engines really speak Chinese (中文), which ones silently do not, and what each one costs you.

Which engines actually speak Chinese

Quick TTS ships four TTS engines and lets you switch between them from one dropdown — but not all four cover every language, and the gaps are not the ones you would guess from the marketing.

Chinese is the one language where this product's usual ranking flips, and pretending otherwise would send readers to an engine that will not appear in their dropdown:

• Supertonic HD does not support Chinese. It is the only one of the sixteen locales left out of the Supertonic language map. The model does not ship Chinese language support, so there is no HD Mandarin tier at any quality setting.
• Kokoro-82M is English-only and hidden outside English.
• Piper's zh_CN-huayan-medium is therefore the only neural Chinese voice in this app — not the budget option, the only option.

The upside is that Piper is the engine with the widest device reach. It runs on WebAssembly rather than WebGPU, which means it works on hardware that cannot run either HD engine — older laptops, machines with GPU acceleration disabled, and (more slowly) mobile. For Mandarin specifically, the engine you can actually run is worth more than the engine with the better spec sheet.

The honest limits

Two honest limits. First, huayan-medium has no smaller fallback voice: the x_low Chinese model in the upstream catalog fails at inference in the WebAssembly runtime we ship, so a device that cannot allocate the medium model falls back to the Web Speech API rather than stepping down to a lighter neural voice. Second, it is one voice — there is no male/female choice in Chinese, which there is in English and Spanish.

There is also a delivery note that matters more for Chinese than any other language. Model weights come from a CDN, and mainland connections to the default hosts are throttled badly enough that a multi-fetch engine init used to fail almost every time. Quick TTS now probes and commits to a domestic mirror for those users, verifies every mirrored byte against a SHA-256 pin, and fetches weights in resumable ranged pieces so a stalled link costs one piece instead of the whole download.

Chinese text is chunked at 120 characters rather than 300 for the same reason as Japanese and Korean — a Han character carries far more speech than a Latin one.

Nothing you paste is uploaded

This is the part that separates in-browser TTS from the cloud tools that look identical in a browser tab. A cloud tool sends your Chinese text to a server, synthesizes it there, and sends audio back — the page is a remote control. Quick TTS ships the engine to you instead and runs it on your CPU or GPU, so your text is never received by anyone. The test is simple: once the page and its model have loaded, the neural engines keep working with your network disconnected.

That matters most for the documents people actually want read aloud — a contract, a medical letter, coursework, an unpublished draft. It also means there is no "we phonemize non-English text on our server" asterisk, which is not true of every tool advertising local synthesis.

Reading Chinese documents, not just pasted text

PDF, DOCX, EPUB, ODT, RTF, HTML, TXT and Markdown are all parsed in the browser, so a whole Chinese EPUB or a long PDF opens and plays without a byte leaving your machine. A scanned PDF — a photo of text with no text layer — is handled by running OCR locally in WebAssembly rather than uploading the scan, which is precisely the case where local processing is worth the most.

Speed, volume and voice can all be changed mid-read and the current passage replays under the new settings without starting over. Switching the engine itself stops playback, so you press play once to start the new one — there is no mid-read handover of the remaining text between engines.

Frequently asked questions

Is Chinese text-to-speech free?

Yes. Quick TTS reads Chinese with no account, no character limit and no watermark. Synthesis runs in your browser on your own device, so there is no per-character server cost to meter and nothing to gate behind a sign-up.

Is my Chinese text uploaded anywhere?

No. The speech engine ships to your browser and runs locally, so the text you paste is turned into audio on your machine and never sent to a server. That is an architectural property, not a promise in a privacy policy — there is no server in the loop once the page has loaded.

What is the best Chinese TTS voice in the browser?

Piper is the only neural Chinese voice here, because Supertonic ships no Chinese support and Kokoro is English-only. It runs on WebAssembly, so it works without a GPU. Open the app at quick-tts.com/zh-cn/ to get it — the neural voice list follows the page language.

Does it work on a phone?

The Web Speech API engine works on every modern phone with no download, and it is the default so the first play click always works. The neural engines are more demanding: Supertonic HD needs WebGPU and is effectively a desktop feature, and Piper runs more widely but can be slow on low-end hardware.

Can I download the Chinese audio?

Yes, as WAV or MP3, generated on your device. Anything longer than about 20,000 characters is packaged as a ZIP of audio parts so memory stays bounded on long documents.

Try it in Chinese — open the 中文 app, not the English one

This is the one piece of setup worth knowing, because it is not obvious: the neural voice list follows the language of the page you are on. On the English homepage the Piper voices are the English ones, and Supertonic reads with the English language tag — so Chinese text pasted there will be pronounced as though it were English. Open the 中文 version instead and the engine picks up Chinese voices, including zh_CN-huayan-medium, automatically.

The one exception is Supertonic's "Auto language" toggle, which switches generation to a language-agnostic mode that reads mixed-language text with the right pronunciation per language without tagging anything by hand. That works from any page — it is the right choice for a document that is part Chinese and part English. For which voice to pick once you are there, the Chinese-language voice guide goes deeper.

For the architecture behind all of this, see how in-browser text-to-speech works; for a tool-by-tool comparison with the other local TTS sites, see the comparison page.