Chinese Text to Speech: Mandarin & Cantonese Voices That Get the Tones Right

Chinese text to speech converts written Chinese — simplified or traditional characters — into spoken Mandarin or Cantonese audio, and the best modern engines get the tones right without any manual markup. That last part matters more in Chinese than in almost any other language: Mandarin distinguishes words by four tones plus a neutral tone, and Cantonese uses six, so a text to speech voice that guesses a tone wrong doesn't just sound foreign — it says a different word. This guide covers how Chinese text to speech engines handle tones and polyphonic characters, the difference between Mandarin and Cantonese voices, simplified versus traditional input, and which tools are worth your time.

How does Chinese text to speech handle tones?

Neural Chinese TTS engines predict tones from context rather than reading characters in isolation. Chinese writing doesn't mark tones the way pinyin does — the character 妈 (mā, "mother") and 骂 (mà, "to scold") are visually distinct, but many characters share a written form and change pronunciation by context. A modern engine runs the sentence through a language model first, decides what each character means in context, and only then generates audio.

Three things separate a good Chinese voice from a bad one:

• Polyphonic characters (多音字). Characters like 行 read *xíng* ("to walk," "OK") in one context and *háng* ("row," "profession") in another. 了 flips between *le* and *liǎo*, 重 between *zhòng* ("heavy") and *chóng* ("again"). Older concatenative voices picked the most common reading and were wrong constantly; neural voices from Azure, Google, and ElevenLabs resolve most of these correctly from sentence context.
• Tone sandhi. Mandarin changes tones in connected speech — two third tones in a row (你好, *nǐ hǎo*) surface as *ní hǎo*, and 不 and 一 shift tone depending on what follows. Engines that apply sandhi rules sound native; engines that don't sound like a beginner reading flashcards.
• Prosody and pauses. Chinese is written without spaces between words, so the engine must segment the sentence itself before it can phrase it naturally. Bad segmentation produces pauses in the middle of words.

If you're comparing tools, paste in a sentence loaded with polyphonic characters and listen. It's the fastest quality test there is.

Mandarin vs Cantonese: pick the right voice, not just "Chinese"

Mandarin (Putonghua) and Cantonese are not accents of one language — they're mutually unintelligible in speech, with different tone systems, vocabulary, and grammar in colloquial writing. A Mandarin voice reading text intended for a Hong Kong audience will be understood in writing-to-speech terms but will sound wrong, and colloquial written Cantonese (with characters like 嘅, 咗, 唔) will trip up a Mandarin voice entirely.

Locale codes to look for:

Locale	What you get	Typical use
zh-CN	Mainland Mandarin, simplified script	Mainland China audiences, most learning content
zh-TW	Taiwanese Mandarin, traditional script	Taiwan audiences; softer, retroflex-light accent
zh-HK / yue	Cantonese, traditional script	Hong Kong, Macau, much of the diaspora

Azure labels its Cantonese voices under zh-HK (e.g., HiuMaan), Google offers Cantonese under yue-HK, and ElevenLabs supports both Mandarin and Cantonese in its multilingual models. Amazon Polly's Zhiyu voice covers Mandarin (zh-CN). If your audience is Cantonese-speaking, confirm the tool has true Cantonese voices — plenty of tools advertise "Chinese" and only ship Mandarin.

Simplified vs traditional input: does it matter?

Yes, but less than people fear. Most major engines accept both scripts and will read either one, because the underlying text normalization maps both to the same phonetic representation. The practical rules:

1. Match script to locale when you can. Feed simplified text to zh-CN voices and traditional text to zh-TW or zh-HK voices. Mixed pairings usually still work, but region-specific vocabulary (软件 vs 軟體 for "software") reads most naturally in its home locale.
2. Watch one-to-many mappings. Converting traditional to simplified is mostly lossless, but the reverse isn't — one simplified character can map to several traditional ones (发 → 發/髮). If you auto-convert scripts before synthesis, use a context-aware converter, not character-by-character replacement.
3. Numbers, dates, and units are normalized by the engine (2026 becomes 二零二六 or 两千零二十六 depending on context). Spot-check anything financial or legal.

Best Chinese text to speech tools

Tool	Mandarin	Cantonese	Strengths	Pricing model
Microsoft Azure Speech	zh-CN, zh-TW	zh-HK	Huge voice catalog, SSML control, pinyin phoneme overrides	Free monthly allowance, then pay-as-you-go per million characters
Google Cloud TTS	zh-CN, zh-TW	yue-HK	Strong neural quality, generous free tier (1M WaveNet characters/month at the time of writing)	Per-million-character pricing (~$4–$16/1M depending on tier)
ElevenLabs	Yes	Yes	Most natural emotional delivery, voice cloning	Free tier (~10k credits/month); paid plans start around $5/month
Amazon Polly	zh-CN (Zhiyu)	No	AWS integration, cheap standard tier	Per-million-character pricing with a 12-month free tier
TTSMaker / free web tools	Yes	Some	Zero cost, no signup	Free with limits

For long-form narration in Mandarin, Azure and Google are the workhorses — predictable pricing and SSML `` tags that let you force a pinyin reading when the engine picks the wrong polyphone. For expressive content (ads, character voices, dubbing), ElevenLabs sounds the most human but costs the most per character. For a broader look at engines across every language, see our text to speech software guide.

How to fix pronunciation when the engine gets it wrong

Even the best engines misread rare polyphones, brand names, and classical Chinese. Your options, in order of effort:

• Rephrase. Often the cheapest fix — swap the ambiguous character for a synonym.
• SSML phoneme tags. Azure and Google both accept pinyin-based phoneme hints, e.g. forcing 行 to *háng* in a banking context (银行). This is the reliable fix for recurring terms.
• Custom lexicons. Azure supports uploadable pronunciation lexicons — worth it if you're producing hours of audio with the same product names.
• Cantonese colloquial text. If a Cantonese voice stumbles on formal written Chinese, try rewriting in written Cantonese (and vice versa); each voice is trained toward one register.

What people actually use Chinese TTS for

The big four use cases we see: language learners generating listening practice with adjustable speed, video creators dubbing content for Chinese-speaking audiences, accessibility (screen-reading Chinese sites for visually impaired users), and businesses localizing IVR or e-learning for mainland, Taiwan, and Hong Kong markets — which is exactly where the zh-CN/zh-TW/zh-HK distinction stops being academic.

TTS also pairs naturally with multilingual publishing: if you're producing written Chinese content for SEO, engines like AI text to speech systems can turn each article into audio for the same audience, and platforms like AutoSEO handle the multilingual article side on autopilot.

If you're building out voices in other languages too, we've covered Japanese text to speech — which shares the kanji-reading ambiguity problem — and Spanish text to speech in companion guides.

Frequently Asked Questions

Can text to speech read both simplified and traditional Chinese?

Yes. All major engines (Azure, Google, ElevenLabs, and most free web tools) accept both scripts. For the most natural results, pair simplified text with zh-CN voices and traditional text with zh-TW or zh-HK voices, since vocabulary and phrasing differ by region even when the characters overlap.

Is Chinese text to speech accurate with tones?

Modern neural voices are highly accurate on tones for everyday text, including tone sandhi (the rules that change tones in connected speech). The remaining failure mode is polyphonic characters — characters with multiple readings — where the engine occasionally picks the wrong one in unusual contexts. Azure and Google let you correct these with SSML phoneme tags using pinyin.

Which tool has the best Cantonese voices?

Azure (zh-HK voices) and Google (yue-HK) both offer solid neural Cantonese, and ElevenLabs supports Cantonese in its multilingual models with the most expressive delivery. Amazon Polly does not currently offer Cantonese. Always test with colloquial written Cantonese if that's what you'll actually be feeding it — some voices handle formal written Chinese far better than 口語 text.

Is there a good free Chinese text to speech option?

Yes. Google Cloud's free tier includes a large monthly character allowance, ElevenLabs has a free tier of around 10,000 credits per month at the time of writing, and browser-based tools like TTSMaker generate Mandarin audio at no cost. Free tiers are fine for testing and short clips; for hours of narration you'll want a pay-as-you-go API for consistency and SSML control.