Audio Tags and Voice Changer Let Users Direct AI Speech Delivery

In Eleven v4, a line of generated speech can be directed with words in square brackets. The tags describe how a passage should sound: an emotion such as nervous or sad, a vocalization such as a cough, or a delivery choice such as whispering. They can be placed anywhere in the script, so a cue can sit beside the words or moment it is meant to shape.

Alec Wilcock starts in Text to Speech, where he types a short voiceover and generates it using his voice clone. The initial read—“Hey, everyone. How’s it going? Today, I want to talk about astrophysics.”—provides a baseline. To make the opening sound nervous, he adds [nervous] before the line and generates it again. In this example, the same words are spoken with a nervous tone.

Tags can be combined to direct different parts of a line. Wilcock’s example pairs a nervous opening with a throat-clearing cue after “Hey, everyone,” then adds [whispering] to shape the word “astrophysics.” As a tag is entered, the interface offers suggestions grouped under Emotion, Vocalizations, and Delivery. The examples shown include angry, bored, tired, and excited; coughing, sighing, and gasping; and whispering, slowly, and softly.

A tag can request an audible action—“Okay, everyone, [coughs] we have some big news”—or set the emotional tone of a sentence, as in “[sad] Oh... and nobody showed up to hear it.” The user can place cues throughout the script rather than rely on a single instruction for the whole generation. Manual tags let the writer choose where each cue appears, then regenerate the line to hear the effect.

Writers who are unsure which tags to use can select Auto-tag after writing a script. Wilcock describes it as scanning the text and adding audio tags based on context. In his example, “I can’t believe you didn’t tell me about this before!” receives a [surprised] tag at the beginning.

He says the feature is especially useful for longer generations and scripts. It offers a way to start with context-based suggestions rather than deciding on and inserting every cue manually. The example shows Auto-tag adding a cue to the line; Wilcock presents it as an aid for longer scripts, not as a replacement for the option to place tags yourself.

Audio tags put delivery cues in the text. Voice Changer uses a different input: a recording of the user speaking the line. Wilcock turns to it when he wants to control a generated voice’s delivery with his own performance rather than rely on tags.

He demonstrates with the line, “Where did I put my glasses? I lose them about once a day,” and selects Rebekah Nemethy’s “Pro Narration” voice. After playing a generated version, he opens Voice Changer, chooses a microphone, records the line, and generates it in the selected voice. The interface also offers the option to upload an audio file. In this demonstration, the output follows the delivery of his recording.

Wilcock describes the converted speech as retaining the original cadence, delivery, and timing. He frames the quality of the recording as important to the result:

Voice Changer’s settings panel shows Stability, Similarity, and Style Exaggeration, as well as Remove Background Noise and Speaker boost. Wilcock suggests trying different values to see what they produce, but recommends leaving the settings at their defaults as a starting point and adjusting from there.