Designing a Voice for an AI Anime Character, Then Adding It Later

CreativeEdge runs a Japanese studio that produces a full AI anime series, and it also publishes a beginner tutorial series built on the knowledge from that production. The third instalment deals with the two pieces that everyone postpones: the main character's voice and the theme song. The studio keeps the whole tutorial inside cheap or free model tiers, which is why the tools in it are ElevenLabs and Suno rather than anything exotic.

The reason this instalment is worth reading is that casting a voice turns out to be the same kind of problem as designing a face. You describe it in writing, you get candidates, you keep one, and then you have to keep it consistent for the rest of the production.

A voice you design once and reuse

In ElevenLabs, Voice Design lets you describe a voice in text, generates three candidates, and saves the one you pick as a named voice. The studio's example is named Jun, and once it exists it can be selected for any line in any scene. One attempt costs 350 credits, which makes the described-versus-generated loop cheap enough to iterate but not free enough to be careless.

The prompt they wrote is the interesting artefact. Translated and condensed, its shape is:

Native Japanese, standard Japanese, adult female, studio quality. Persona: lively, neutral fantasy character. Emotion: energetic, curious, impulsive. Very high pitch, light and thin, neutral. Suppress chest resonance and vocal thickness. Bright and compact, slightly nasal in colour. Clear articulation, fast syllable transitions. Fast tempo, short pauses, quick reactions, wide pitch swings. Not a mature feminine voice, not breathy and refined, not warm and rich, not deeply resonant. Not natural, calm adult speech, but the quick, exaggerated delivery of a fictional character.

Two halves are doing different work. The first half specifies the target. The second half forbids the thing the model would otherwise produce. Left alone, a voice model defaults to the pleasant, warm, well-modulated narrator: chest resonance, controlled pacing, a hint of breath. That voice is pleasant everywhere and wrong for an anime lead, and you cannot prompt your way out of it by describing the right voice harder. You have to name the wrong one and take it off the table. Anyone who has fought a specific image style out of an image model will recognise the move.

The prompt also asks for a voice that is deliberately unnatural as human speech. The studio wants an animated, theatrical register, not a person talking. That is a target worth naming out loud: stylisation is a requirement, not a defect, and models will not volunteer it.

The wall on child voices

The tutorial then hits a policy boundary that matters for anyone building a cast, and it is worth separating the rules from the experience.

The Prohibited Use Policy, which the studio cites as of its 17 August 2026 update, bars submitting voice recordings of anyone under 18 to create or train a voice model, and bars content designed to be recognised as a specific real minor.

ElevenLabs' help documentation draws a second line that is narrower and more practical: children's voices, and adult voices designed to sound child-like, cannot be added to the Voice Library at all. That restriction covers synthetic voices as explicitly as recorded ones.

The studio then reports the outcome from its own production, which is blunter: the boy's voice could not be produced, and the studio says plainly that it should not be attempted. Treat that as a tested result rather than a written rule, because the documentation is about the library and what a creator actually hits is the practical wall.

The workaround in the series is the one animation has used for decades: cast an adult voice and place it. Their example voice is an adult woman, pitched high and thin, with the resonance and warmth explicitly stripped out. Anime has a long tradition of adult performers voicing young characters, and it exists for the same reason here. The restriction does not remove the casting choice, it narrows which performer reads for the role.

If you are planning a cast with several young characters, that constraint belongs in the plan and not in the week you record dialogue.

Sync is a trade-off you get to choose

ElevenLabs' text-to-speech has its own failure mode, and the studio's answer is blunt: use the newest model. At the time of writing that is Eleven v4. Older versions mispronounce Japanese and get intonation wrong often enough to be a problem, and the current model resolves most of it. Voice quality on a non-English language is not a stable property across versions, so pin the version you validated and re-check when it changes.

Then comes the part that shapes the whole pipeline. Kling 3.0 Omni can lip-sync to a supplied audio track. In the Firefly workflow the studio uses there is no equivalent step for a generated clip, so it produces silent video and adds the voice afterwards in a simple editor, which avoids paying for another service just to attach audio. Firefly does ship a lip-sync option inside its Translate Video feature, but that synchronises a translated reading of speech already in the clip and is limited to enterprise customers with eligible plans, which is a different job from accepting an arbitrary track.

They compared the two audio paths and reached a conclusion that sounds wrong until you look at the pictures. Kling's built-in voice generation is natural and immersive, but it mispronounces a lot and its accuracy degrades past about ten seconds, which drags the mouth off the audio. ElevenLabs reads cleanly, and the mouth does not match it exactly. For a stylised anime, that mismatch is acceptable. For live action it would not be. The tolerance is a property of the art form, not of the tool, and it is why this order of operations works in animation and would fail in a talking-head video.

There is a practical bonus buried in that choice. Audio added in post is a separate layer. You can rewrite a line, re-read it, change the delivery, or swap the language without touching a single generated frame. Audio baked into a generation is a new render. If you want the opposite approach, where the mouth is rebuilt to match the audio instead of the other way round, that exists too: Amazon's dubbing pipeline redraws an actor's lips frame by frame, at the cost of treating every frame as an edit.

The theme song, and why the expansion matters

For the opening theme, the studio uses Suno in its simple mode with a one-line prompt: an early-1980s Japanese adventure anime theme, sung by a chorus of boys and girls. Suno writes the lyrics and expands the prompt, and the expanded version is the thing worth reading, because it shows what the model decided you meant:

1980s Japanese TV anime theme song, adventure, bouncy march-like pop rock, bright synth brass and strings, sparkling arpeggios, strong live drums and driving bass, unified mixed choir of boys and girls, straightforward and sunny delivery, harmonies opening up in the chorus, short synth intro into the vocal, thickening rhythm section and strings through the verses, a final chorus with layered choir, clear melody, warm analog-style mix.

None of that came from the one-line prompt, and all of it is editable. The expansion is your real prompt: it tells you which defaults the model filled in, so you can correct the ones you disagree with instead of re-rolling and hoping. It is the audio equivalent of reading a model's rewritten prompt before you accept the image.

Two housekeeping details the studio flags. Downloads are metered per plan and each one has to be unlocked, so decide what you need before you spend the counter. And the free tiers are for experiments, not for anything you intend to publish commercially. If you only need a background track rather than a titled theme song, a general-purpose model may be enough; we tested one of those in our look at Lyria 3.5 inside Gemini.

The order of operations

The transferable lesson is in the sequencing rather than the tools. Cast the voice as a design task, early, with the same care you give a character sheet. Expect the casting pool to be constrained by policy, and plan around it rather than discovering it during production. Then keep dialogue as a post-production layer, because that is the choice that lets you fix a performance after the picture is locked. The lip-sync drift that comes with doing it the other way is not a bug you can prompt away.