Controlled voice tests expose a sharp accent gap
Artificial Analysis finds that 12 of 49 speech models perform significantly differently across US and UK English using identical cloned voices.
The benchmark
Artificial Analysis has introduced a Controlled Voice Arena that evaluates speech models with the same eight cloned reference voices: four from the United States and four from the United Kingdom. The design is intended to isolate model quality from the familiar problem that one provider’s default voice may simply be more appealing than another’s.
The initial results show that accent changes can materially alter rankings. Artificial Analysis says 12 of 49 tested models have statistically significant performance differences between US and UK English, with non-overlapping confidence intervals. Qwen-Audio-3.1-TTS-Plus leads the US-English comparison, while HeyGen Voice performs better in UK English and leads the overall table. The result is not a universal victory for either system; it is a reminder that speech quality is partly regional.
Why it matters for voice agents
Voice systems are increasingly used for customer service, education, accessibility and entertainment, where an accent mismatch can affect comprehension and perceived trust. A model that sounds excellent to American listeners may be less convincing to British listeners even when it receives the same text, prompt and reference voice. That makes a single global leaderboard a poor proxy for deployment quality.
The controlled setup is a useful methodological step because it separates the voice being cloned from the model generating the performance. It also gives buyers a more realistic question to ask: how does this system behave for the actual accents, markets and users it will serve?
The benchmark remains limited. Eight English voices cannot represent the world’s linguistic diversity, and preference-based Elo scores capture listener judgments rather than every production concern such as latency, pronunciation consistency, safety controls or licensing. Still, the release pushes accent coverage toward a first-class procurement metric. For companies building voice agents across regions, “best voice model” is becoming an incomplete question.
Uncle Cat take
Qwen’s US lead and HeyGen’s UK advantage show that accent robustness is not a cosmetic detail; vendors selling one global quality score are hiding deployment variance.