Fish Audio leads Hume's voice cloning leaderboard

MarkTechPost compared seven voice cloning API providers using a single reference recording: 10 seconds of one speaker in a quiet room, saved as WAV. Every platform read the same two sentences through its instant cloning path with default settings. Prices were checked on September 20, 2026.

Hume published its Voice Replication Leaderboard on September 10, 2026, testing 11 models across 25 reference voices and 7 prompts. Three blind raters scored each clip from 1 to 5 on how closely it matched the reference. Among the providers covered, Fish Audio s2-pro led with 4.03. Cartesia sonic-3.5 scored 3.70, ElevenLabs Multilingual v2 3.68, Cartesia sonic-3.6-beta 3.63 and Inworld TTS-2 3.62. ElevenLabs Eleven v3 came last of the 11 models at 2.91.

The same board shows why a single score can mislead. Cartesia sonic-3.6-beta topped naturalness at 4.36 but ranked 8th on identity, while Inworld TTS-2 led audio quality at 4.61.

Pricing and consent gates

ElevenLabs recommends 1 to 2 minutes of clean audio for Instant Voice Cloning; Professional Voice Cloning needs at least 30 minutes and adds the strictest consent step in the group, a Voice Captcha where the voice owner reads on-screen text aloud. Eleven v3 and Multilingual v2 cost $0.10 per 1K characters; Flash, Turbo and v3 Conversational cost $0.05.

Inworld clones from 3 seconds and charges nothing for cloning itself, with TTS-2 starting at $25 per 1M characters and falling to $12.50 at $1,500 a month. Fish Audio bills $15 per 1M UTF-8 bytes, which roughly triples for Chinese, Japanese and Korean text. Cartesia bills 1 credit per character, from $5 for 100K credits to $299 for 8M. Gradium, founded by Kyutai co-founders, clones from 10 seconds but covers only English, French, German, Spanish and Portuguese. Resemble AI requires explicit, verifiable consent from the voice talent and offers PerTh watermarking on every output.

The full comparison is available at MarkTechPost.