Chinese AI Voices

Ten voices, and the order they arrive in

Ten distinct voices covers what an ordinary scene needs and stops short of a crowd. Casting inside a project will not hand two characters the same voice until all ten are spoken for, so an ensemble separates itself and you do not have to police it — the constraint only bites on Plus and Pro, since Free caps a project at two cast members and Basic at five.

The order matters because each library character is fixed to one of the ten. When a second character in the project wants a voice that is taken, the one added later moves to an unused one — among voices written for its own gender first, then the shared one — and characters already in the project keep what they have, so adding someone in week six does not invalidate the audio you have already rendered. What separates two characters on the same voice is the performance: age, temperament and the direction on each line, which is real but is a smaller difference than a change of voice. So build the cast list before you audition, and listen to the characters that had to move rather than trusting the description.

多音字, and respelling with a homophone

Polyphonic characters are ordinary in Chinese and concentrated exactly where fiction needs precision: in names. 单 is shàn as a surname and dān otherwise; 解 is xiè as a surname and jiě otherwise; 仇, 覃, 区 and 乐 behave the same way. Outside names, 行 and 重 and 长 change reading with sense, and a wrong pick is a wrong word rather than an odd inflection.

There is no reading-override list, and Chinese has no phonetic spelling equivalent to Japanese kana. What it does have is homophones: replace the misread character with one that has the same syllable, the same tone and only one reading, and the listener hears the name you meant — write 善 for the surname 单 and it comes out shàn. The cost is in the subtitles: the SRT export on Plus and Pro follows your script text, so the stand-in character appears there and has to be changed back in your editor for a subtitled release. Audition every name before you render an episode, because that is where the problem shows.

Tone sandhi breaks at the seams you create

Third-tone sandhi, 一 shifting tone with what follows, 不 before a fourth tone — inside a phrase these usually come out right. Where it fails is at boundaries you introduce: a line break or a comma dropped into the middle of a word stops the sandhi from applying, and 你好 read as two separate syllables is immediately audible to a native listener even though nothing is misspelled.

So the practical rule for Mandarin scripts is the opposite of the advice for most languages: do not split lines for pacing. Keep each utterance whole and let the line-level speed control on Plus and Pro handle the pacing instead, because that changes delivery without changing where the engine thinks the words end.

A scene, cast line by line

Six lines in a workshop, three speakers, all of them characters from this page. The Mandarin is the script; the notes after each line are what we would cast and how we would direct it. Someone has not turned up, and somebody is going to be blamed for it.

Kestrel: 单师傅说他不来了。 — Master Shan says he isn't coming. The surname 单 is why this line is first: it is shàn as a surname and dān everywhere else, and nothing in the sentence tells the engine which you meant. Audition it before anything else in the script. If it comes out dān, write 善 instead and the listener hears the right name.

Orrin: 他不来，这锅谁背？ — He isn't coming, so who carries this? 背 is a polyphone too, bèi or bēi, and here it is bēi. Orrin is the old man on the porch and the register fits. Neutral, no preset.

Koji: 我背。反正我也没别的可背的。 — I'll carry it. I've nothing else to carry anyway. 我也 is a pair of third tones, and the first shifts to a second tone when the two are said together — but only if they stay in the same line. A comma or a line break dropped between them stops it, and a native listener hears the difference even though nothing is misspelled. Koji is the clipped rival; mark this one sad if you have line-by-line emotion, which is a Pro control.

Kestrel: 你别抢。抢完了还得你收拾。 — Don't grab it. If you do, you're the one clearing up. Two sentences in one line, which is right. Keep utterances whole in Mandarin and use line-by-line speed on Plus and Pro for pacing instead of breaking them apart.

Orrin: 让他抢。年轻人抢一回，能记十年。 — Let him. A young man grabs once and remembers it for ten years. 一 changes tone with what follows, and 一回 is the case to listen to here. Neutral: the line is a grandfather's verdict, and a preset would turn it into a speech.

Koji: 那我记着了。 — Then I'll remember it. Four characters, and the whole turn of the scene. If it lands flat, regenerate — a second take of a short Mandarin line differs enough to be worth the second try.

Three speakers, three of the ten base voices, standard Mandarin throughout, because regional accents are not something you can ask for. The work in this scene was not the casting. It was one surname, one polyphone, one tone-sandhi pair and one place where the line must not be broken.

What to listen for before you export

Exporting spends download credits and re-exporting the same project spends them again, so listen once to the assembled scene. Six things catch Mandarin out more often than the rest.

Names, first. Every surname and every invented name, auditioned before anything long is rendered: 单, 解, 仇, 覃, 区, 乐, and anything out of a xianxia glossary. Where a reading is wrong, swap in a character with the same syllable, the same tone and only one reading — and remember the swap shows in the SRT subtitles on Plus and Pro.

Polyphones outside names. 行, 重, 长, 都, 还, 为. Listen to the sentence rather than the word: most are right in context, and the ones that are not tend to sit inside four-character phrases, where the context is compressed to nothing.

Numbers. 一 and 不 shift tone with what follows, 二 and 两 are not interchangeable, and digit strings use 幺. Listen to every number that is spoken and write it the way it is said: 两点半, 幺三八.

Where you broke the lines. Play the mixed file and listen for sandhi that did not happen. A comma dropped into the middle of a phrase is the usual cause; put the phrase back together and use line-by-line speed for the pacing you wanted.

Latin letters and mixed text. App names, English abbreviations and single letters inside a Chinese line come out either as English or as letters. Decide which you want and write it that way; where a Chinese reading is what you need, write the Chinese.

Long narration. Mandarin narration past about forty characters flattens. Listen for where the intonation stops moving, and split at a full stop rather than inside a phrase, so that nothing loses its sandhi on the way out.