Intel / post

HeyGen Voice takes the top spot on a blind TTS leaderboard

Source: x.com
Date: Published Oct 9, 2026

From the Daily Brief

HeyGen released HeyGen Voice, its own text-to-speech model, and made it available in the product and through its API. Artificial Analysis independently places it first in its Controlled Voice TTS Arena with an Elo of 1,201, ahead of Qwen and ElevenLabs models in blind comparisons. The pricing needs a careful read. Artificial Analysis lists HeyGen Voice at $30 per million characters, compared with $19.3 for Qwen and $40 for ElevenLabs v4 Turbo, while HeyGen is offering a launch price of $15 per million characters through October 31. The tradeoff is speed: Artificial Analysis measures generation at about 40 characters per second, slower than some rivals. For voice agents where perceived quality matters more than latency, it is worth a benchmark during the discounted window. For realtime conversation, the generation speed may rule it out.

Read the 2026-10-10 Brief →

Context

When HeyGen says its new text-to-speech model is "#1 in voice", the board it links to in the launch thread is Artificial Analysis's Controlled Voice Arena leaderboard. On that board every model is compared using the same eight cloned voices, four US English and four UK English, so the comparison is between models reading with identical voices. Artificial Analysis publishes a separate Provider Voice leaderboard as well; HeyGen's claim refers to the controlled one. Scores are Elo ratings, a rating system built from head-to-head results, anchored so that the OpenAudio S1 model sits at exactly 1000. HeyGen's replies describe the underlying contests as blind tests in which listeners pick the clip they prefer.

The lead is real on the page but narrow. When the leaderboard was captured, HeyGen Voice was rated 1199, plus or minus 15, from 1,938 samples, a 95% interval of 1184 to 1214. Second-placed Qwen-Audio-3.1-TTS-Plus was rated 1185, with an interval of 1171 to 1199, and Eleven v4 and Eleven v4 Turbo followed, tied at 1158. Because the two intervals overlap, the board's own rank column lists HeyGen Voice as possibly first or second rather than clearly first. It also has fewer samples than any other model in the top ten. Artificial Analysis's own launch-day post reported different figures, an Elo of 1,201 across 1,468 arena appearances against 1,182 for the Qwen model, so the numbers on the live board have moved since the announcement.

The board covers English only, and when people in the replies asked about other languages HeyGen said it would share more later. On cost, the leaderboard lists HeyGen Voice at $30 per million characters on HeyGen's API, against $19.3 for the Qwen model, $40 for Eleven v4 Turbo and $80 for Eleven v4. HeyGen says the model is live in its app and API, and that API usage is automatically billed at $15 per million characters through October 31.

Read the source x.com