Create a custom voice from a text description, then use its voice ID in TTS and speech to speech.
“A soft-spoken man in his late sixties from Glasgow, Scotland, gentle and slow, with a slight rasp. Perfect broadcast quality audio.”
Morning, love. The kettle's on, and I've found the old photographs you were asking about.
Now this one is your grandmother at the seaside, the summer before I met her.
“A natural, age-appropriate nine-year-old girl from Toronto, curious and bright, speaking at a quick pace with real excitement.”
Did you know octopuses have three hearts? I read it in my library book this morning, and I told everyone at lunch.
Okay, next question. Which planet has the most moons? I think it's Saturn, but check for me.
“A warm, upbeat woman in her late twenties from Lagos, Nigeria, quick and confident, with a smile in her voice.”
Hi Tunde, your rider is two minutes away. He has the jollof rice and the extra plantain you asked for.
Your refund is done. It should show in your account by tomorrow morning.
“A middle-aged woman from Mexico City, a late-night radio host, low and warm, speaking slowly and close to the microphone.”
Buenas noches, Ciudad de México. Son las once en punto, y esta canción es para quien todavía va camino a casa.In English: Good evening, Mexico City. It's eleven o'clock, and this song is for anyone still on their way home.
Gracias por quedarte conmigo esta noche. Mañana volvemos a la misma hora.In English: Thanks for staying with me tonight. We are back tomorrow at the same time.
“A calm young man in his twenties from Tokyo, soft and measured, guiding a slow breathing exercise before sleep.”
ゆっくり息を吸って、四つ数えます。そして、静かに吐いてください。In English: Breathe in slowly and count to four. Then breathe out quietly.
今日も一日、おつかれさまでした。肩の力を抜いて、目を閉じましょう。In English: You made it through another day. Relax your shoulders and close your eyes.
“A fast-talking cab driver in his fifties from Brooklyn, New York, chatty and a little gruff, always in the middle of a story.”
So I tell the guy, buddy, the bridge is closed, we're taking the tunnel, and he looks at me like I stole his lunch.
Alright, we're here. Watch the puddle when you get out, it's deeper than it looks.
Each voice was designed from the description on its card, with no audio input. Line 1 is the preview the design call returned. Line 2 is a new line read by the saved voice ID on Realtime TTS-2.
Describe a voice, get three in under 5 seconds. Publish one and use it everywhere, just as fast.
POST /voices/v1/voices:design
{
"designPrompt": "A soft-spoken man in his late sixties from Glasgow, Scotland, gentle and slow, with a slight rasp. Perfect broadcast quality audio.",
"previewText": "Morning, love. The kettle's on, and I've found the old photographs you were asking about.",
"languageCode": "en-GB",
"voiceDesignConfig": {
"numberOfSamples": 3
}
}Name the age range, the city or region for the accent, the pitch and pace, the tone and the texture. The locale in languageCode sets the accent, so en-GB and en-NG give different voices from the same words. Write the script the way this person would talk, because the model shapes the voice to fit it.
{
"languageCode": "en-GB",
"previewVoices": [
{
"voiceId": "1framework__design-voice-8bd78905",
"previewText": "Morning, love. The kettle's …",
"previewAudio": "<base64 WAV>"
},
{
"voiceId": "1framework__design-voice-f1457c58",
"previewText": "Morning, love. The kettle's …",
"previewAudio": "<base64 WAV>"
},
{
"voiceId": "1framework__design-voice-db03c38c",
"previewText": "Morning, love. The kettle's …",
"previewAudio": "<base64 WAV>"
}
]
}{
"displayName": "Site mock glasgow",
"tags": [
"site-mock-2026-10-11"
]
}Response{
"voiceId": "1framework__design-voice-8bd78905",
"source": "TVD",
"languageCode": "en-GB"
}“Now this one is your grandmother at the seaside, the summer before I met her.”
{
"text": "Your refund is done. It should show in your account by tomorrow morning.",
"voiceId": "1framework__design-voice-d521d405",
"modelId": "inworld-tts-2",
"audioConfig": {
"audioEncoding": "MP3"
}
}{
"type": "session.update",
"session": {
"type": "realtime",
"model": "inworld/models/gemma-4-31b-it",
"audio": {
"output": {
"model": "inworld-tts-2",
"voice": "1framework__design-voice-d521d405"
}
}
}
}The voice agent answered“I can check that for you; could you please provide your order number?”
“A gentle grandmother with a soft Irish lilt, slow and cozy, reading a bedtime story by the fire.”
Request 1And when the last light went out, the little fox curled up and dreamed of the sea.
Request 2And when the last light went out, the little fox curled up and dreamed of the sea.
POST /tts/v1/voice
{
"text": "And when the last light went out, the little fox curled up and dreamed of the sea.",
"modelId": "inworld-tts-2",
"voiceDesign": {
"designPrompt": "A gentle grandmother with a soft Irish lilt, slow and cozy, reading a bedtime story by the fire."
},
"audioConfig": {
"audioEncoding": "MP3"
}
}
A grandfather, described in one sentence
Write it in English. Descriptions under 30 characters rarely give a useful voice, and conflicting words such as fast-paced and slow make results less consistent.
The preview text must make 1 to 30 seconds of audio, about 50 to 400 characters in English. Write it in the language you want the voice to speak.
Set languageCode to a locale such as en-GB, en-NG or es-MX. Leave it out to detect the language from the script.
By default the description is turned into a structured voice profile for you. Send DESIGN_PROMPT_MODE_VERBATIM to write the profile yourself, one attribute per line.
Voice design is included on every plan, from free. A saved voice costs what a library voice costs: about $1.50 an hour of speech on demand and $0.75 on Growth ($25 and $12.50 per million Realtime TTS-2 characters, at about 1,000 characters a minute). The same key calls LLM Inference, with open models from $0.07 per 1M input tokens.
Each saved voice takes one custom voice slot. The free plan has 100, Creator 500, Builder 3,000, Developer 10,000 and Growth 30,000.
Design voices, not impersonations. Impersonating a real person without consent, or hiding that people are talking to AI, breaks our Acceptable Use Policy.
The Acceptable Use Policy prohibits impersonating another person or organization without their consent or the legal right to do so. The policy covers every input you send and every output you create, so it covers designed voices too.
An app that speaks with a designed voice has to tell people they are interacting with AI, unless that is obvious from the context.
Text descriptions count as input under our service terms. You may not send input or create output that you lack the rights for, or that violates anyone's rights.
To copy a voice that already exists, use voice cloning, which asks you to confirm you have the rights to that voice before it runs.
Ask your agent to add Inworld.
curl -s https://api.inworld.ai/voices/v1/voices:design \
-H "Authorization: Basic $INWORLD_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"designPrompt": "A warm, upbeat woman in her late twenties from Lagos, Nigeria, quick and confident, with a smile in her voice.",
"previewText": "Hi Tunde, your rider is two minutes away. He has the jollof rice and the extra plantain you asked for.",
"languageCode": "en-NG",
"voiceDesignConfig": {"numberOfSamples": 3}
}' > previews.json
# three draft voices, each with a WAV preview
jq -r '.previewVoices[].voiceId' previews.json
for i in 0 1 2; do
jq -r ".previewVoices[$i].previewAudio" previews.json | base64 --decode > preview_$i.wav
doneThree previews, each a draft voice ID with a WAV preview. Wrote previews.json and preview_0.wav to preview_2.wav.
AI voices built for people run on Inworld. Who are you building for?
Inworld AI voice design creates a new voice from a written description, while voice cloning reproduces a real voice from an audio sample you already have. Use cloning when you have recordings and the rights to them, and design when you need a voice that does not exist yet. Both give you a voice ID that works in every text to speech request.
A good description for Inworld AI voice design names the gender and age range, the accent with its city or region, the pitch and pace, the tone and the texture, such as "a soft-spoken man in his late sixties from Glasgow, gentle and slow, with a slight rasp". Write the script the way that person would talk, because the model shapes the voice to fit it. Descriptions can run up to 1,000 characters of English, and ones under 30 characters rarely give a useful voice.
Yes, Inworld AI voice design returns up to three previews per request, and you publish the one you want to keep as a permanent voice ID. Previews are drafts, and text to speech does not accept a draft voice ID, so call the publish endpoint with the preview's voice ID. The voice is saved to your library under that same ID and works in every TTS request like a library voice. The Growth plan holds up to 30,000 custom voices, and the free plan 100.
Yes, a voice published with Inworld AI voice design is a standard voice ID, so you pass it as the output voice in the speech to speech session configuration. It is no slower than a library voice. A designed voice reached its first audio chunk in a median of 259 ms, against 264 ms for a library voice.
A voice made with Inworld AI voice design can speak any locale in the text to speech language list, such as en-GB, en-NG, es-MX or ja-JP, which you set through languageCode. The description itself is written in English, and the locale carries the accent. Write the script in the language you want the voice to speak, and on Realtime TTS-2 the voice can read text in about 500 languages and dialects.
No, Inworld AI's Acceptable Use Policy prohibits using voice design to impersonate another person or organization without consent or the legal right to do so. To reproduce a real voice with permission, use voice cloning, which asks you to confirm you have the rights to the voice. For a new voice, describe the traits you want, such as age, accent, pace and tone, and the model makes a voice to match.
Yes, voices you make with Inworld AI voice design can be used in production, including commercial applications, under the Inworld terms of service and subject to the Acceptable Use Policy. Every plan includes a commercial license. Voice design is included on every plan too, and a saved voice bills per character at the same rate as a library voice, so a designed voice costs no more to use.
Yes, Inworld AI voice design is available on every plan through the Voices API, the playground and the SDKs. Ad-hoc design inside a TTS request, with voiceDesign.designPrompt, is a preview feature and may change before it is marked stable. If you need the same voice every time, publish it and use its voice ID.