Create speech and reuse a voice

Design, clone and reuse a voice, then check the words in the actual audio.

Updated 2026-10-22

Give a character a voice

1. Start a conversation and describe who is speaking: language and accent, gender, age, persona, emotion, then timbre and pace.
2. Paste the exact words to be spoken, and spell names and unusual words the way they should sound.
3. If you want to choose the voice yourself, say so in the brief. Otherwise the agent casts it for you.
4. Send the brief. The agent designs the voice, which returns several short auditions.
5. The agent saves the audition that fits and records your lines with the saved voice.
6. Play the recordings in the conversation and ask for changes to any line.

Design a warm, unhurried narrator voice for our bakery radio spot: American English, woman in her forties, friendly and confident. Let me hear the auditions before you record. Script: "Fresh bread every morning at Marlow Street Bakery. Open from seven."

Check that it worked

1. Play every recording from start to finish with sound.
2. Check each required word, names and pronunciation against your script.
3. Listen for clipped line endings, long silences and level jumps between lines.
4. Ask the agent to transcribe the recording and compare it with your script. Generated speech sometimes drops or garbles a word, and a file that plays is not proof it says the right thing.

If a line is wrong, name it and ask for that line again. Renders vary, so a second take often fixes it; a line that keeps failing usually needs rewording.

Design a new voice

A voice design turns a written description into several short auditions. Each audition is kept as its own audio file in the conversation.

• Description: the agent turns your brief into a voice description of 20 to 1,000 characters. It works best in this order: language and region, gender, age, quality, persona, emotion, then timbre and pacing.
• Audition script: when your lines total 100 to 1,000 characters, the auditions can read them. With shorter lines, the auditions read a sample written for the voice instead.
• An audition is not the performance. It is a sample. Your lines are recorded only after one audition is saved as a voice.

Hear the auditions yourself

By default the agent picks the audition that best fits your brief. It can compare from two to six candidate voices against your brief and script before it chooses. If you want to choose, ask in your brief to hear the auditions first and end with "Wait for my go-ahead". Leave Step by step off for this: the web app can't approve its steps yet. See Step by step.

Clone a voice from recordings

Upload recordings of one speaker and ask the agent to clone the voice. The clone becomes a saved voice that can read any new script.

• A video or a song as the source: the agent can take the audio out of a video clip and separate the voice from music first. Speech separation needs a recording of at least 4.6 seconds.
• Background noise: noise removal is an explicit choice for each clone. Ask for it when your recordings are noisy; cleaning an already clean recording can lower quality.
• Too little audio: a clone with less than 14 seconds in total is refused before anything is sent to ElevenLabs, and nothing is charged for it. The refusal states how many seconds your recordings measure. Add more recordings of the same speaker and ask again.

Record speech with a voice

Once a voice is saved, the agent records your exact lines with it.

• One voice per character. The agent keeps one voice for each character, so the same person sounds the same in every scene.
• Several speakers. A scene with more than one speaker can be recorded as one dialogue, with each line given its own voice.
• Delivery. Describe pace, emotion and emphasis separately from who the voice is. For a shouted or held line, say so in the brief.
• Talking on camera. A character's own words over their visible face need a take made with the speech. See Video, speech, music and finishing.

Choose the speech model

The agent chooses the speech model unless you pick one. To pick, open the quality menu beside Send, choose Show model choices, and set the Voice row. ElevenLabs is the default for new conversations. The row also lists other speech models, such as OpenAI speech and MiniMax speech; one your workspace cannot run is marked Unavailable.

A model you pick applies from the next message you send. The agent cannot swap in a different speech model: a step that needs another model is refused, not substituted. Designed and cloned voices are ElevenLabs voices, so leave the Voice row on ElevenLabs or Auto when you want to use one.

Reuse a saved voice

A saved voice keeps its identity, so it can read new scripts in later conversations.

1. Open a new conversation.
2. Choose Add from Library and select a recording made with the voice (or the voice itself, if it is in your Library), or name it in your brief, for example "use the narrator voice from our spring launch conversation".
3. Paste the new lines and send.

Adding a voice to a brief does not generate anything; sending the brief starts new speech. Only a saved voice can be reused. An unsaved audition is a preview and cannot be attached as a reusable voice.

What voice work costs

Voice work is charged in credits like other media. Use Check what this message costs under Planning budget in the quality menu beside Send to see an estimate before you send.

Work that is refused before it runs returns its whole reservation. See Credits.

If something goes wrong

• A design or clone was interrupted. Wavemaker does not repeat a paid voice design or clone on its own. A clone or save whose outcome is uncertain is looked up, not created a second time, so you do not end up with duplicate voices. Reopen the conversation and check the latest state before asking again. See Recover interrupted work.
• The words are wrong. Ask for a transcript of the recording against your script and re-record only the lines that differ.
• The voice sounds different between lines. Ask the agent to use the one saved voice for that character and re-record the lines that drifted.
• A recording won't play or download. Refreshing a link does not regenerate the audio. The original file stays in the conversation and your Library.

Related guidance

• Understand credits, reservations and charges
• Library: find, organize and reuse your work
• Recover interrupted or uncertain work
• Use references, characters and styles
• Video, speech, music and finishing