Japanese AI Voices

Pitch accent, and the names it gets wrong

Pitch accent in Japanese is phonemic. 箸 and 橋 are the same two morae with the fall in a different place, and getting it wrong does not produce an accented reading, it produces a different word. For ordinary vocabulary it is usually right. Where it is not dependable is proper nouns, and fiction is mostly proper nouns.

Kanji readings make it worse: 東 is Azuma or Higashi or Tō depending on whose name it is, and an invented character name has no dictionary entry at all. There is no reading-override list, so the working method is to write the name in kana in the script wherever the engine gets it wrong. Kana is unambiguous, and it is one place where Japanese is easier to fix than Chinese. The cost shows up in the subtitles: the SRT export on Plus and Pro is built from your script text, so the name appears there in kana, and a subtitled release means changing it back in your editor.

Politeness lives in the text, not in the casting

です・ます and plain form are the same character speaking, and the shift between them is usually a story event — the moment two people stop being strangers is written in verb endings, not in tone of voice. It follows that the register lives in your script, and that you should keep the same voice across the shift. Recasting the character at the moment they drop keigo destroys exactly the beat you were building.

The same applies to the first-person pronouns. 私 / 僕 / 俺 / あたし tell a Japanese listener more about a character in one word than a paragraph of description would, and they are a writing decision that no voice setting can stand in for. Pick them before you cast, because a boku character and an ore character want different voices.

Ten voices, and no children

Ten distinct voices is enough for most drama-CD-shaped scripts, where the speaking cast in any one scene is rarely large even when the series cast is. It is not enough for a class-of-thirty ensemble scene. Basic caps a project at five cast members, so in practice the ceiling is something only Plus and Pro projects reach. Ten is not a Japanese limitation — it is the same ten everywhere — and each character is fixed to one of them and separated from the others by how it is directed: age, temperament, the way it holds a pause. Two characters sharing one voice are still distinguishable, but the difference is smaller than a change of voice, so put your two leads in first and let the crowd double up.

The harder limit is age. There are no children's voices in Japanese, and none in any other language here. A child character is performed by an adult voice playing young, the way animation has always done it — often a woman's voice for a boy. It reads as a performance of a child rather than as a nine-year-old, and no amount of direction closes that gap. For anime-shaped work, where child characters are routine, this is the constraint to plan around: write the young roles for teenagers, keep them off-mic, or listen to one on your own lines before you commit to a series.

A scene, cast line by line

Six lines in a club room, three speakers, all of them characters from this page. The Japanese is the script; the notes after each line are what we would cast and how we would direct it. Koji is quitting, Ezra is the one who asks, Kestrel is the adult in the room.

Koji：先に言っておく。俺は明日、部活をやめる。 — I'm telling you first: tomorrow I'm leaving the club. 俺 is the casting brief, and it rules out half the library before you audition anyone. Koji is written as the clipped rival and fits. Neutral, not angry: the decision has already been made, and anger would turn a statement into a threat.

Ezra：……理由、聞いてもいい？ — Can I ask why? Ezra's two-second pause before honesty is exactly this line. The leading ellipsis is read as a pause rather than as a word, which is worth confirming once on your own text. Sad is the preset if you have line-by-line emotion, which is a Pro control; otherwise leave him neutral, because the same preset would be wrong on his last line.

Koji：聞いても変わらない。 — Asking won't change it. Short, plain form, nothing softening it. Neutral, and on Plus or Pro take it to 0.9×: the slowest line in a Japanese scene is usually the one that means it.

Kestrel：やめるのは構いません。ただ、今日の片づけまではお願いします。 — Leaving is fine. Just stay through today's clean-up, please. This is the register change the page is about: Kestrel speaks です・ます to a student, and that single fact tells a Japanese listener her position without a word of description. Keep her neutral, and do not recast her warmer to soften it — the politeness is the softening.

Koji：……はい。 — Yes. One word, and the scene turns on it, because he has answered in the form he uses with her rather than the one he uses with Ezra. Nothing in the casting layer can mark that; it is in the text. If it does not land, regenerate rather than redirect — a second take of a one-word line differs far more than a second take of a long one.

Ezra：今の「はい」、さっきより素直だったな。 — That yes was more honest than the last one. Happy, very lightly. Listen to whether the quotation marks around はい come out as a small lift or as nothing at all; either is fine, and if they are read aloud as punctuation, remove them and the line still works.

Three speakers, three of the ten base voices, and nothing in the setup that says accent — the performance is ordinary standard Japanese and is not something you can steer. What separated these three was pronoun, verb ending and pace, all of which you wrote before anyone was cast.

What to listen for before you export

Exporting spends download credits and re-exporting the same project spends them again, so the listening pass is worth doing once on the assembled scene. Six things go wrong in Japanese more often than the rest.

固有名詞, first and three times each. Every name, place and invented term. A wrong kanji reading is a different word, not an accent, and there is no reading list to correct. Where one will not come right, write it in kana in the script: kana is unambiguous, and the kana will appear in the SRT subtitles on Plus and Pro.

Pitch accent pairs. 箸 and 橋, 雨 and 飴, 神 and 髪. You are listening for where the fall lands, not for the vowels. Ordinary vocabulary is usually right; where the word is short and carries the whole line, take a second take before you accept the first.

Counters and numbers. 一人, 二日, 十日, 四人, 八分. The readings are irregular and a digit gives nothing away. Listen to every counted thing in the script, and write the reading in kana — ひとり, ふつか, とおか — wherever it comes out wrong.

Speech-level drift. Play the whole export and listen for a character sliding between です・ます and plain form where you did not intend it. It is almost always a typo in the script rather than a performance problem, and the script is where it gets fixed.

Katakana English. ウイルス, レベル, and any brand-shaped coinage. Listen for whether the word comes out Japanese or English; one English-sounding word inside a Japanese scene is louder than it looks on the page. Writing it in katakana, with the vowels you want, is how you decide that yourself.

Long lines. Japanese lines past about fifty characters flatten in the middle, the same as every other language here. Listen for where the intonation stops moving and split into two lines for the same speaker, which adds a short gap rather than a full stop.