AI Voice and Music

Synthetic voices got good enough in 2024 that many listeners cannot reliably pick them out of a lineup. That is useful for clean narration at midnight. It is dangerous when the voice belongs to somebody who never said the words.

Kinds of synthetic voice

Text to speech

Voice Studio runs on ElevenLabs. Its Text2Speech tool reads text in a chosen voice (English and Spanish), with controls for stability, style, speed and pitch, and exports MP3, WAV or OGG. It can also add a voiceover with burned-in captions to a video. Speech2Speech changes a recording into a different voice while keeping the original timing and pacing. Speech2Text transcribes. SoundEffects makes effects up to 22 seconds from a prompt. The page also lists AudioIsolation and MusicGeneration. All six tools draw on one pool of ElevenLabs credits, and the page warns in capital letters that they are expensive: a few minutes of voice can use up a monthly quota.

Synthetic voices stumble on the same things human readers do, only more confidently:

• Numbers: "38%" may come out as "thirty-eight percent" or "three eight percent." Spelled-out text ("thirty-eight percent," "twenty twenty-six") is read as written.
• Abbreviations: "IPUC" is a coin flip. "The I-P-U-C" or "the Public Utilities Commission" is not.
• Breath: punctuation is the only breathing guide. A 40-word sentence in a synthetic voice sounds like a hostage reading a statement.
• Stray words: synthetic voices occasionally add a word, drop a word or mispronounce a town name, so every take needs a listen before anyone relies on it.

Voice design

Voice design makes a new voice from a description. ElevenLabs' docs recommend being specific about age, gender, accent, tone, pacing and emotion or persona. Each generation returns three options, and one gets saved to a voice slot. The docs also advise a full sentence or short paragraph of preview text, since short phrases sound abrupt.

A description like "sounds like [famous person]" is a clone by the back door, and providers block it for good reason.

Cloning and consent

ElevenLabs offers Instant Voice Cloning from about 1 to 2 minutes of clean audio and Professional Voice Cloning from about 30 minutes to 3 hours. For the professional version, its docs say only your own voice can be cloned, verified by reading lines into your microphone, and that someone else's consent does not change this. A colleague who wants to share her voice makes and verifies the clone on her own account and shares it.

Even where a tool allows instant clones of other people, the standard this course holds is simple.

A consent note does not need legal language. It needs specifics, like this one:

I, Marisol Vega, agree that the Harlan Valley Library Friends
may create a synthetic copy of my voice using ElevenLabs,
to narrate three explainer videos about the library levy,
published on the group's website and social accounts
between November 1, 2026 and May 31, 2027.
The group will show me each script before it is voiced.
I can withdraw this consent in writing at any time;
the group will then stop making new audio and remove
published audio within 14 days.
Signed / dated

The U.S. Copyright Office's 2024 report on digital replicas found real gaps in federal and state protection against unauthorized voice and likeness copies and recommended a new federal law. As of October 2026 a federal bill, the NO FAKES Act, had cleared the Senate Judiciary Committee but was not law. Lesson 7 covers what is law.

Multi-voice scenes in Dialogue Editor

Dialogue Editor turns a script with speaker names into one audio conversation. The script is pasted or loaded, each speaker gets a voice from a browser filtered by gender, age, accent and category, and the "Eleven v3 (dialogue, expressive)" model generates it. Stability has three labeled settings at 0, 0.5 and 1, running from Creative (most varied) through Natural to the steadiest delivery. There is also a Seed field. The result downloads as MP3.

It suits dramatized scenarios that are clearly invented: "a resident calls the water office," a two-voice FAQ, a radio-style explainer. Recreating a real conversation between real people, even from a transcript, is a different thing: a reenactment with fake voices. At minimum it must be labeled as one. Usually it should not be made.

Music in MusicStudio

MusicStudio makes full songs on Mureka. It takes up to 5,000 characters of lyrics with section tags like [Verse] and [Chorus], a style prompt built from style chips (tempo, key, time signature and more), and a vocal gender. An "Instrumental (no vocals)" mode makes a bed. Each generation returns 1, 2 or 3 songs, and the page says each returned song is billed separately. Voice Studio's MusicGeneration tool is another route for instrumental beds.

A typical narration-bed prompt is instrumental, simple and steady:

Instrumental, acoustic guitar and soft upright bass, light brushed drums, 92 BPM, warm and steady, no big builds, no vocals.

The bed is then ducked under the voice. The module on audio production covers this in Music beds and ducking, and SoundStudio's theory page covers cleaning up an AI-generated song at #t-suno and loudness at #t-loud.

Licensing quick check

• ElevenLabs: its help center says the free plan has no commercial license, and published free-plan content must credit elevenlabs.io or 11.ai in the title (Eleven Music in place of that for music). Paid plans include a commercial license, except beta services.
• Mureka (MusicStudio): the MusicStudio page does not state licensing terms. Mureka's current terms for each plan decide whether a song can go in anything paid or sponsored.
• Suno and Udio: both were sued by major labels in 2024. Universal settled with Udio in October 2025, and Warner settled with Suno in November 2025, with Suno moving to licensed models in 2026 and limiting downloads on free accounts. Terms changed as a result, so old assumptions about downloads and ownership may be wrong.

Disclosure

Honest disclosure says what is synthetic, where a listener can find it. YouTube's policy requires disclosure for AI music that is the main focus of a video (its examples list "AI generated music" without that qualifier), while cloning your own voice for a voiceover does not. Even where disclosure is not required, a line costs nothing. Typical examples:

• End card: "Narration: AI voice (ElevenLabs). Music: AI-generated (Mureka)."
• Podcast notes: "This episode's dramatized calls use synthetic voices. The people are invented."
• Robocalls are a separate world: the FCC ruled in February 2024 that AI voices count as "artificial" under the Telephone Consumer Protection Act, so robocalls using them need the called party's prior consent. Any campaign or group putting an AI voice on a phone dialer is covered by that ruling.