AI Text to Speech

Text to speech, or TTS, turns written text into spoken audio. Older TTS read words one after another in a flat voice. Modern AI text to speech reads the whole passage first, so it understands which word to stress, where a sentence turns into a question and when a pause makes a point, and it speaks with the breath and rhythm of a person.

On Mage, text to speech runs on Eleven Speech, ElevenLabs' latest model, Eleven v4. You paste up to 5,000 characters, pick one of 16 voices, and get an MP3 that lasts exactly as long as the reading. Audio tags in square brackets, such as [whispers], [laughs] or [excited], direct the delivery without being read aloud.

It lives in the same account as Mage's image and video models. A narration can go straight under a video, onto a character as its voice, or into the Lipsync app to make a portrait speak it.

Text to speech reads exactly what you give it, so write for the ear. Spell out what should be said the way it should sound: "twelve hours" rather than "12h", "the year twenty twenty-six" if the number should be read as a year, and full words for abbreviations that are not meant to be spelled out. Break long sentences into shorter ones; a listener cannot scroll back.

Punctuation is the cheapest way to direct pacing. A comma is a short breath, a full stop a longer one, an ellipsis a hesitation, and a dash a sudden break. A question mark changes the melody of the whole sentence, so check that questions really end with one.

• Write numbers as words when the reading matters
• Keep sentences short and spoken
• Use ellipses for pauses and dashes for breaks
• Read the script aloud once before you generate
• Split very long texts into chapters

The 16 voices cover the most common jobs. For long narration and audiobooks, pick a steady storyteller such as Darian, Florence or Sawyer. For news, training and explainers, a crisp anchor such as Alicia or Eldrin. For ads and social video, an upbeat, modern voice such as Elowen, Jade or Lawrence. For games, a character voice such as Kaelen.

Every voice speaks every supported language, so a Spanish or Japanese script works with any of them. Generate the first two or three sentences in a few voices, compare them, and then run the whole script in the one that fits.

Eleven v4 reads tags in square brackets as direction. [excited], [nervous], [sarcastic] and [softly] change the emotion; [whispers] and [shouting] change the volume; [sighs], [laughs] and [gulps] add a reaction; [door creaks] or [applause] add a sound inside the reading. Put a tag right before the sentence it changes, and use one or two per sentence.

Tags make dialogue and character lines come alive, while plain narration often needs none. For character voices that must stay the same in every video, see the AI voice generator guide, and for songs and music beds, the AI music generator.

The same voices run through the Mage API: send the text and a voice to POST /v1/eleven_speech/generate and get an MP3 back, priced in blocks of 250 characters. The Mage API page covers keys and pricing. With the Mage MCP server connected, Claude, ChatGPT and other assistants can read a script aloud for you in the chat, after quoting the Gems it costs.