Artificial Analysis finds US and UK English accent gaps in TTS and speech models

Artificial Analysis reports that 12 of 49 models show statistically significant performance differences between US and UK English, with non-overlapping confidence intervals.

Seven models perform better in US English and five favor UK English, and the average gap across all models is 7 Elo points.

In its TTS comparison, Lightning V3.1 Pro scores 94 Elo points higher in US English than in UK English, the largest gap among the models listed. Qwen-Audio-3.1-TTS-Plus scores 66 points higher in US English, while Soniox TTS Real-Time v2 and Speech 2.8 Turbo score 51 and 49 points higher in UK English. The firm says TTS models should be judged in the accent relevant to the use case rather than by overall rankings alone.

Artificial Analysis says its Controlled Voice Arena uses the same eight cloned voices for every model, four US English and four UK English, with separate leaderboards for each accent.