Kokoro TTS vs ElevenLabs: one documentary script, two AI narrators
Data as of 2026-11-10. Rechecked every quarter.
Kokoro is an open-weight text-to-speech model with 82 million parameters. It is free, Apache-licensed and small enough to run on an ordinary CPU, which is why people looking for an ElevenLabs alternative keep finding it. We make short history documentaries, so we tested both on the job we actually have: one 112-second script about Project Pigeon, B.F. Skinner’s wartime plan to let pigeons steer a guided bomb, narrated twice over the same pictures. Kokoro ran on our own desktop Mac with its af_heart voice. ElevenLabs ran through its API with the eleven_v4_turbo model and the Olivia voice, one request per shot. Listen to the first two shots of each below. Both read every word correctly. The difference is delivery. To our ears Kokoro was a clear reading of a script, while ElevenLabs was hard to tell from a human narrator. The one tell we noticed is that it never hesitates: a person reading this script would drop in an “uh” or an “um” now and then. Writing a few of those into the script, or using ElevenLabs’ audio tags, might close that gap; we have not tried it yet. Kokoro also stopped for more than a second between sentences, against about half a second for ElevenLabs, which made the Kokoro cut feel slower. ElevenLabs takes audio tags: in another film of ours we wrote “The emus? [amazed] They’re still there.” and got surprise where the story needed it; Kokoro has no audio tags, so the way to change its delivery is to pick another voice or speed. We kept the ElevenLabs read for the finished film. The cost was small: 1,554 characters, about $0.06 on our ElevenLabs Starter plan, against $0 for Kokoro. Speed was close: ElevenLabs returned the whole film in about a minute, and Kokoro took about half the film’s length on the Mac’s CPU. Our workflow now uses both: Kokoro while the script and pictures are still changing, because every rerun is free, and ElevenLabs once, for the final cut. We cache each shot, so after an edit we only pay again for the lines that changed. If you have no budget at all, Kokoro is good enough to publish a clear explainer; if the voice is the product, as it is in a documentary, the paid read was worth the six cents.
1. The facts side by side
| Tool | Pricing | Free tier | Visits/month | Revenue (30 days) | Agent-ready | Free credits | Tested |
|---|---|---|---|---|---|---|---|
| ElevenLabs elevenlabs.io | — | — | 33.6M (2026-09) | — | llms.txt, MCP | Referral code | eleven_v4_turbo, Olivia voice, one API request per shot with timestamps: 1,554 characters for a 112-second film, about $0.06 on the Starter plan, about a minute for the whole film. Sounded like a narrator, and audio tags such as [amazed] changed the delivery where we placed them. We used it for the final cut. (2026-11-02) |
| Kokoro Web kokoroweb.app | free | — | — | — | none found | — | We ran the same Kokoro 82M model locally with its Python package (af_heart voice), not through this app. Free, about half the film’s length to generate on a desktop Mac CPU, every word read correctly, but a read-aloud delivery to our ears and pauses of more than a second between sentences. No audio tags. We use it for drafts. (2026-11-02) |
“—” means we have not looked or the site does not say. Prices are as the tool’s own page states them.