Inworld TTS 2 for character voice

Inworld TTS 2 is the row for a character read. You describe the delivery in natural language — tired, close to the mic, a mascot who is too pleased with a pun — and you get a wav. 100+ languages. 5 credits. It is not a talking generate. If we see the mouth, a lipsync row still has to wear this file.

Search Console still shows "inworld tts-2" (11 impressions, position 9, seven days to 4 September 2026). The sibling page, Inworld TTS 2: a voice file, not a talking generate (5cr), is the catalog boundary. This page is when the job is a character: a mascot, a game NPC, a series narrator with a personality, a localised persona that has to stay the same person in Korean and in French.

A character is a performance note, not a face

Inworld TTS 2 is Inworld Realtime TTS-2 (Research Preview): most powerful and expressive in that roster, 100+ languages, natural-language style steering for fine-grained delivery control. Category: text-to-audio only. No image required. No duration chips. No resolution. Display price: 5 credits. Featured. The output is speech.

Style steering is the character lever. "Tired, close to the mic, almost a whisper" is a performance note. "Bright, slightly too fast, like a Saturday-morning mascot who knows the slogan is silly" is a performance note. "Older, dry, unimpressed, then one warm line on the product name" is a performance note. None of those are lighting setups. None of them turn a product still into a presenter. Audio tags and style notes punch the take. They do not track lips.

That is why this row wins when the audience will know the voice across clips. A series with a recurring narrator. A game character who has more lines than you will ever film. A UGC-style ad where the persona is audio-first and the picture is a pack shot or b-roll. Kapwing's 2026 UGC figures (digest §2.8, 4 September 2026) are about creative volume; they do not require a generated face on every take. A locked still plus a character wav plus captions is often the honest stack.

The catalog does not list voice-clone on this slug. Cloning is a different Inworld row if you need a likeness of a real person. TTS 2 is text-to-audio: script in, speech out. Do not assume research-preview power includes a founder clone. If the brief is "our actual CEO," record them or clone on the row that actually clones, with consent.

100+ languages is why a character can travel. A French take and a Korean take are two wavs of the same performance note, not two faces. They are not one lipsync pass that "just switches language." You generate the reads, then you decide whether a mouth has to match.

The door is AI text-to-speech. Ranked speech sits on best text-to-speech model. This page is not a ranking. It is the job: character read.

Write the read like a director, then chunk it

A wall of paragraph will flatten on TTS 2 the same as on any other engine. Chunk it.

• One beat per generate when the character turns: joke, then sincere close.
• Put the style note on every chunk, not only the first. The model does not remember yesterday's take unless you say it again. That is AI asset versioning: keep character_style.txt and paste it.
• Mark pauses in the script if the row honours them; do not expect a 40-second sermon to breathe because you feel it should.
• Product names: spell them as they should be said. All-caps SKUs get read as letters. Write "Nura Glow" how you want it heard.

Style notes that work are sensory and social, not brand adjectives.

Works: "Half-whisper, indoor, slightly rushed, as if the room next door might hear." Does not: "Premium, luxury, on-brand, Gen-Z but classy."

Works: "Deadpan. The joke is that she will not perform the joke. Warmth only on the last three words." Does not: "Make it pop."

Works: "Small creature voice, not a child, not a parody, quick consonants, delighted at the cap click." Does not: "Cute."

Research Preview is in the description. Treat the row as a powerful expressive engine with that label attached, not as a finished actor-replacement contract. 5 credits is the display price (per 1,000 characters on the matrix). Plan the speech job at 5 credits. Plan the mouth separately.

If the character is on-camera, stop before you call the job done. Generate the line here — 100+ languages is why you might — then drive the face on a lipsync model. Kling Avatar Pro is the image-to-lipsync shape (still required, 4K, 58-credit floor). Kling Lipsync is the video-to-lipsync shape (7 credits) when you already have footage. TTS 2 does not open a jaw.

If the character is off-camera, you are done when the wav lands. Lay it under I2V product motion, under b-roll, under a screen recording. Caption the speech for mute feeds. Do not generate a host "so it feels like a video."

When this is the wrong character tool

You need a specific human. Clone or record. TTS 2 is not that slug.

You need a song. Suno or a music row. Speech is not a chorus. Firefly audio is not Suno is the other confusion; this one is speech versus picture.

You need the face and the voice in one click. Talking generate: Veo 3.1, Seedance, another native-audio video row. You will get an invented person unless you locked a reference the row accepts. That person will not be "the Inworld character" you spent a week voicing.

You need 15 languages of a talking head by Friday. Budget two rows per language: TTS 2 (or a clone) plus lipsync. 100+ languages is the speech coverage, not a mouth coverage.

The script does not fit the slot. Time it first on the script timing estimator (in-browser, no credits). 150 wpm is a typical VO target. Cut the script before you buy the voice.

Keep the character bible short enough to paste: age impression, social attitude, distance to mic, one example line that sounds right, languages in use, whether we ever see teeth. Store it next to the wavs. When a take wins, do not "improve" the bible in place; branch it.

Gemini 3.1 Flash TTS and Cartesia Sonic 3.5 are other speech rows in the catalog. Use them when they win a blind listen. Inworld TTS 2 is the one whose published hook is 100+ languages plus style steering. That pair is what a character-localisation job actually eats.

The test

Listen with the picture off. If you can cast the voice in one sentence — "that's the mascot," "that's the tired expert" — the row did the character job.

Unmute the cut. If we see a closed mouth, you still owe a lipsync or you should have kept the face off-screen.

If the same persona has to recur next week, you need the style note on disk, not a memory of a slider. Five credits will not remember.

A character read is audio. Inworld TTS 2 writes that audio. Everything that looks like a person talking is a later, different generate.

Write the line, paste the style note, generate the wav from text-to-speech. Sign up when the persona is decided and the next file should be speech, not a new face.

FAQ

Does Inworld TTS 2 output a talking character on screen?

No. Text-to-audio, 5 credits, no image, no resolution. You get a voice file. A visible mouth is a lipsync or talking-head row that takes a plate.

How do I keep the same character across languages?

Reuse the same style note. Generate a new wav per language. Do not expect one lipsync pass to switch language by itself. If the face is on screen, run lipsync per wav.

Is style steering the same as cloning my founder?

No. Style steering is a performance description on this research-preview TTS row. Cloning a real person is a different Inworld slug, with consent. TTS 2 will not become a likeness because you wrote a name in the style box.