Cekura tests 26 speech models and finds speed does not buy naturalness
Tarush Agarwal's benchmark puts Inworld TTS-2 Flash first for full-sentence latency, while Fish Audio Drama 3 leads an automated naturalness estimate.
By Ryan Merket · Published
Primary source: X
Why it matters
Voice-agent builders face tradeoffs among response time, pronunciation and interruption handling. Cekura's benchmark makes those choices easier to inspect, while its role as both evaluator and vendor gives buyers reason to examine the methodology alongside the leaderboard.
Cekura published a public benchmark of 26 text-to-speech models on October 19th, comparing latency, reading accuracy, naturalness and interruption behavior for voice agents. The test's fastest model began speaking after a median 46 milliseconds; the model with the highest estimated naturalness ranked well behind it on speed, according to the results shared by co-founder and CEO Tarush Agarwal (@tarush_agarwal_).…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.