AI Voice Design: Create a New Voice From a Text Prompt

AI voice design creates a new synthetic voice from a written description of how it should sound, with no recording of a real person needed. You describe the speaker, the tool returns candidate voices, and you save the one you like as a reusable voice. It differs from voice cloning, which copies a real speaker's voice from audio.

How AI voice design works

Most tools follow the same pattern, though the details differ. The steps below describe the common flow, based on the published workflows of several vendors:

Write a voice description covering the speaker, tone and delivery.
Add sample text for the voice to read, usually one to a few sentences.
Generate one or more candidate voices, typically one to three per request.
Listen to each candidate, then save the one you want as a reusable voice ID.

Saving matters. Google's Gemini API returns a permanent voice ID for a prompted voice, and Inworld publishes a chosen voice under the same voice ID. Once you have a saved voice, you can reuse it without repeating the design step.

Voice design vs voice cloning

Both produce a custom voice, but they start from different inputs. Cloning needs audio of an existing speaker, and design needs only words.

| | AI voice design | AI voice cloning | |---|---|---| | Starting point | A written description of the voice | Audio of a real speaker | | Recording needed | No | Yes, from a few seconds to hours depending on the tool | | Result | A new voice that matches the description | A copy of a specific person's voice | | Consent questions | Fewer, since no real voice is copied | Consent required from the speaker | | Repeatability | Save the chosen voice and reuse it | Save the clone and reuse it |

Because design does not copy a real person, it raises fewer consent questions than cloning does. It does not guarantee that a designed voice is unique, though. No source reviewed says that designed voices never resemble real people. For the cloning side, read what voice cloning is and how to clone your voice.

How to write a voice description prompt

Good prompts cover several dimensions. ElevenLabs' guidance lists language, gender, age range, quality level, a short persona, emotions, timbre, pacing and delivery. The list below combines that advice with the other vendors' notes:

Age: use specific terms, such as "in their 20s," rather than vague ones.
Gender and persona: a short role or persona helps the model settle on one character.
Accent or region: name the accent and where it is from. ElevenLabs says "thick" often works better than "strong" for accents.
Timbre: describe texture separately from emotion, for example warm, low or breathy.
Pace and pitch: state the speed you want, such as a moderate, unhurried pace.
Audio quality: a phrase such as "studio quality" is a common cue.
Context: Hume recommends describing the setting, role or intent, such as a customer support call.

Here is an example prompt built from those dimensions: "A woman in her 20s with a warm, clear timbre, a calm and friendly tone, and a moderate pace, speaking as a customer support agent in studio quality." Write one or two clear sentences, and avoid contradictions.

Voice design tools compared

The table below summarises the candidate counts and prompt rules that each vendor documented as of October 2026. Vendors change these pages, so verify the current details before you build on them.

| Tool | Candidates per request | Prompt guidance or limits | Persistence | |---|---|---|---| | ElevenLabs Voice Design | 3 | Granular prompts; charged by preview-text characters | Save to a voice slot | | Hume Octave 1 | 3 | Prompt plus sample text; no emojis or markup | Voice library, referenced by name or ID | | Google Gemini API | Preview returned with the voice | One or two clear sentences | Permanent voice ID; 200 per project; 1-year TTL | | Inworld | Up to 3 | Description up to 1,000 characters; write in English | Publish to library | | Qwen3-TTS VoiceDesign (open source) | 1 per generation | Natural-language instruction | Reuse the clip as a clone reference | | Parler-TTS (open source) | 1 per generation | Text description of gender, pitch and style | Reuse the description or a speaker name |

Google's voice design documentation uses the v1beta endpoint, so check its status before you depend on it in production. The Hume documentation may also be dated, so confirm current model support before you build.

ElevenLabs' tool generates three voice options from a prompt, and you are charged once based on the characters in the preview text, not per sample. Its guidance says a longer preview text tends to produce more stable and expressive results. Inworld notes that a description under 30 characters rarely gives a useful voice.

Open-source options

Two open models let you run design yourself. Qwen3-TTS's VoiceDesign model takes target text and a natural-language description, and the project is released under Apache-2.0. A designed clip can serve as a reference for cloning, so the voice can be reused across lines.

Parler-TTS, released by Hugging Face under Apache-2.0, is steered by a text description of gender, pitch and speaking style. Its project page gives the example "A female speaker delivers a slightly expressive and animated speech with a moderate speed and pitch." Including a speaker name such as "Jon" in the description gives more consistent voices, according to the project.

Common mistakes to avoid

Contradictory terms. Pairing "fast-paced" with "slow" confuses the model. Pick one pace.
Effect words. ElevenLabs advises against words such as "reverb," "echo" or "phone," which can change the sound in unexpected ways.
Prompts that are too short. Inworld says a description under 30 characters rarely gives a useful voice.
Markup in the prompt. Hume's guidance is to avoid emojis, HTML tags and Markdown.
Mixing emotion into timbre. Describe the sound of the voice, then the feeling for the line.

How OndaLabs voice design works

OndaLabs custom voices can be designed from a text description, as well as cloned from a recording you have permission to use. A voice design or clone costs 1,500 credits. The plans list custom voice slots: 3 on Creator at $22 per month, 10 on Pro at $49, and 30 on Business at $199. Commercial use is included on Starter and above.

Voice design is separate from delivery direction. In OndaLabs, you control delivery by describing it in plain words and with supported expressive tags in the script. Onda Flash is the most expressive model and is available on paid plans at 1.5 credits per character. Onda Lite is available on all plans at 1 credit per character. To try a custom voice, open the custom voices page. For credit and plan details, see the pricing page.

Frequently asked questions

What is AI voice design?

AI voice design is the creation of a new synthetic voice from a written description. You provide a description and sample text, the tool generates candidate voices, and you save one for reuse. No recording of a real person is needed.

How is voice design different from voice cloning?

Voice design starts from a description, while voice cloning starts from audio of a real speaker. A cloned voice copies a specific person, so it needs consent. A designed voice is a new voice that matches the description you wrote.

How do I write a good voice description prompt?

Cover age, gender or persona, accent, timbre, pace, emotion and audio quality. Keep the description to a few clear sentences, and avoid contradictions and effect words such as "reverb" or "echo." Test two or three versions and keep the one that sounds closest to your goal.

Can I make the same designed voice again?

Yes, if you save it. Google returns a permanent voice ID for a prompted voice, and Inworld publishes a voice under the same ID. Saving the chosen candidate and reusing it is more reliable than regenerating from the same prompt, because results can vary with wording.

Can I use a designed voice commercially?

Check each vendor's terms. None of the sources reviewed says that every designed voice is licensable for commercial use. For OndaLabs, commercial use is included on Starter and above, as stated on the pricing page.