Direct a line of dialogue with Eleven v4’s voice controls

In an audio story, a quiet welcome can make a visitor’s arrival feel more unsettling than a shout. The words may stay the same while the performance changes what we think is happening. That is a useful small task for a synthetic voice: explore two readings of one line.

ElevenLabs launched Eleven v4 on 28 September with natural-language direction and audio tags written into the script. It is available through the company’s creative tools and API. This exercise uses the website’s Text to Speech playground, following documentation; no audio was generated or assessed for this article.[1]

A free account includes 10,000 credits a month for practice across ElevenLabs’ products. For commercial work, generate the recording during a paid subscription: ElevenLabs says recordings made on the free plan cannot be used commercially. Paying later does not change the terms under which an earlier recording was made.[5][6]

Keep the speaker fixed

Open Text to Speech, choose a voice from your available Voices and select Eleven v4. Start with a voice whose ordinary delivery suits the character. Keep that voice and model selected throughout the exercise so you are comparing performances from the same starting point.[4]

Paste that text, or a short passage of your own, into the input box and generate a first reading. Listen for the meaning already present in the voice: welcoming, suspicious, relieved or something else. Keep this first version as a reference before adding direction.[4]

Eleven v4 has two voice settings. Stability controls variation in delivery: lower values permit more expression, while higher values stay closer to a fixed baseline. Similarity controls adherence to the reference voice, with higher values potentially reducing naturalness. Leave them unchanged for the first comparison.[2]

Give the line a different intention

For the second reading, place [whispers] before the same words. This documented tag directs vocal delivery. Compare the result with the first reading: are the words still intelligible, does the voice remain recognisable, and has the quietness changed the scene in the way you intended?[3]

If the line needs curiosity rather than secrecy, replace the whisper instruction with [curious]. Avoid stacking several conflicting directions while finding the performance. ElevenLabs’ guidance recommends explicit voice descriptions because a vague tag can sometimes be interpreted as a sound-effect request. The direction is a request for a performance, not a guaranteed result.[3]

Now try one change to the wording’s rhythm: “Come in… There’s something I need to show you.” The added pause gives the invitation a different shape. ElevenLabs documents ellipses as a way to introduce pauses and capitals as a way to add emphasis; ordinary punctuation supplies rhythm. Keep the rest of the passage steady while listening to that change.[3]

If successive takes vary too widely, raise Stability and try again. If the performance feels too fixed, try lowering it. Eleven v4 does not offer the Speed or Style sliders described for some other models, and it does not support SSML. For this model, work with the script, audio tags and its two supported voice settings.[2]

Changing the text and generating again uses credits. The website can offer free regenerations under specified conditions, and its button indicates whether the next one is free. Check the displayed cost before each new version.[7]

Save the take, with its direction

To retain a reading, open History in the right-hand panel of Text to Speech. On a narrow screen, use the history icon above Generate speech. Select the recording’s download icon and choose MP3 or WAV. Keep the baseline and accepted revision under distinct names.[4]

If you share a free-plan recording non-commercially, ElevenLabs requires elevenlabs.io or 11.ai in its title. Paid-plan commercial use still requires the necessary rights and compliance with the applicable terms; Beta Services are excluded from that permission.[6]

Record the voice, model, settings and exact script beside the files. Listen to the chosen clip in its intended place in the story. The best reading for the exercise is the one that makes the visitor’s arrival mean what you want it to mean. A more dramatic delivery is only useful if the scene needs it.

Report an error or suggest a correction