Paste the script. Choose how it is read.

Forty-six named voices across three families. ElevenLabs has twenty: Aria, which is the default, plus Roger, Sarah, Laura, Charlie, George, Callum, River, Liam, Charlotte, Alice, Matilda, Will, Jessica, Eric, Chris, Brian, Daniel, Lily and Bill. Minimax has seventeen: Wise Woman, the default, plus Friendly Person, Inspirational Girl, Deep Voice Man, Calm Woman, Casual Guy, Lively Girl, Patient Man, Young Knight, Determined Man, Lovely Girl, Decent Boy, Imposing Manner, Elegant Man, Abbess, Sweet Girl 2 and Exuberant Girl. Qwen has nine: Vivian, the default, plus Serena, Uncle Fu, Dylan, Eric, Ryan, Aiden, Ono Anna and Sohee. Cartesia offers exactly one voice and no picker.

Nine. ElevenLabs v3 is the most expressive and reads emotion out of the writing; Multilingual v2 holds one voice across thirty languages; Turbo v2.5 is half the price and most of the quality, which makes it the draft option. Minimax Speech 2.8 Turbo gives you emotion, pitch and volume as real controls rather than hints, and 2.8 HD is the same voices at higher fidelity. Qwen3 TTS 1.7B covers eleven languages with automatic detection and 0.6B is the smaller, faster one with the same voices. Cartesia Sonic 3.5 finishes while you wait. Minimax Voice Clone reads your script in a voice from a recording you supply.

Per 1000-character block, at the rate of the model you choose, with a one-block floor — a single short line still costs one block, so it is worth batching short lines into one run where the script allows. ElevenLabs Turbo v2.5 is the cheapest of the ElevenLabs three at about half the price. Failed runs are refunded.

Yes, on every family. Speed runs from 0.7 to 1.2 on ElevenLabs, Qwen and Cartesia, and from 0.5 to 2 on Minimax. The narrow band is narrow deliberately: pushed further, the read stops sounding like a person. Before reaching for it, note that these models read punctuation as direction — a comma is a breath and a full stop is a beat, so write the pauses in rather than asking for them.

It is an ElevenLabs control running from 0 to 1, defaulting to 0.5. Low values give a more expressive read that varies between takes; high values give a consistent, flatter one. Similarity, at 0.75 by default, is the separate control for how closely the render holds to the reference voice, and style, at 0 by default, pushes the delivery away from neutral. None of the three appears on the other families.

On Minimax, yes — it takes one of seven named emotions: neutral, happy, sad, angry, fearful, disgusted or surprised, alongside a pitch shift of −12 to +12 semitones. Both are Minimax-only. On ElevenLabs the equivalent handle is the style control, from 0 to 1.

Up to 5000 characters, enforced as you type. Longer scripts need splitting into runs — break at a paragraph rather than mid-sentence so the phrasing holds across the seam.

Qwen3 TTS covers eleven with automatic detection: Auto, English, Chinese, Spanish, French, German, Italian, Japanese, Korean, Portuguese and Russian. ElevenLabs Multilingual v2 is the other multilingual option, holding one voice across thirty languages.