See the top models on the Artificial Analysis Speech leaderboard: Vote for model

Artificial Analysis compares speech generators using shared reference voices so that a listener's favorite voice does not dominate the comparison. Its English Controlled Voice leaderboard uses the same eight cloned voices, four American and four British, for every model.

Listeners vote between pairs of recordings with model names hidden, and the votes are converted into Elo-style preference ratings with uncertainty intervals. This helps compare how models deliver speech in a similar voice; it does not turn the rating into a percentage of words pronounced correctly. The evaluator measures pronunciation in a separate benchmark described in its methodology.