Best Voice Cloning APIs in 2026: Speaker Similarity, Consent Checks, and Price per 1M Characters
MarkTechPost Asif Razzaq
Voice cloning APIs were ranked by how closely they mimic one speaker. Fish Audio led on similarity, while ElevenLabs had the toughest consent checks.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
A voice clone has a higher bar than a decent synthetic voice: it has to sound like one specific person. MarkTechPost tested seven cloning providers with the same 10-second reference clip, then compared them on how much audio they need, how they check consent, what commercial use costs, and how many languages they cover.
The similarity results came from Hume’s Voice Replication Leaderboard, published on September 10, 2026. Fish Audio’s s2-pro came out ahead among the providers in this roundup, with a 4.03 score. Cartesia’s sonic-3.5 followed at 3.70, ElevenLabs Multilingual v2 at 3.68, Cartesia sonic-3.6-beta at 3.63, and Inworld TTS-2 at 3.62. ElevenLabs’ Eleven v3 landed last of the 11 models tested, at 2.91. And the board also showed why one number can fool you: Cartesia sonic-3.6-beta was strongest on naturalness, while Inworld TTS-2 topped audio quality.
Consent is where the field really splits. ElevenLabs uses rights attestation for Instant Voice Cloning and Voice Captcha for Professional Voice Cloning, which means the owner has to read on-screen text aloud. Fish Audio requires live ownership verification for pro clones. Resemble AI asks for explicit, verifiable consent, and Hume says Octave uploads must come from a consenting speaker. The other services lean more on terms and attestations than on technical checks.
The audio requirements are all over the map too. Inworld can work from 3 seconds for instant clones, while Cartesia uses 10 seconds and Gradium also starts at 10 seconds. Hume needs 15 seconds, so the same 10-second test clip falls short there. Professional cloning gets much heavier: ElevenLabs wants at least 30 minutes and ideally 2 to 3 hours, Cartesia’s pro clones start at 30 minutes, and Inworld’s beta professional cloning needs 10 minutes.
Pricing mirrors the split between fast demos and serious deployments. Inworld’s TTS-2 can fall to $12.50 per 1M characters, Fish Audio charges $15 per 1M UTF-8 bytes, and ElevenLabs’ Eleven v3 sits at $0.10 per 1K characters, or $100 per 1M. Hume’s Creator tier starts at $14 a month, while Gradium, Cartesia, and Resemble all sit in different bands depending on plan and use case. The neat summary is simple: nobody wins every axis, and the cheapest option is not the one with the best clone.
My take — AI-written commentary, not fact-checked reporting
This is what happens when a market gets serious: the flashy demo stops mattering as much as consent, access, and the boring math of pricing. The uncomfortable bit is that the best-known name is not the clean winner on similarity, which is exactly the sort of thing vendor marketing hopes no one checks.
Read more about this at: MarkTechPost