Inworld TTS-2 Flash Tops Cekura Speed Test

This is a meaningful result for voice-agent builders because it pairs near-instant startup with strong pronunciation accuracy. The caveat: pipeline latency, interruption handling, and edge-case errors still matter more than a single leaderboard number.

• –46 ms median latency makes first-response delay nearly invisible in many voice-agent turns.
• –Cekura separates full-text, streaming, and first-audio measurements, so vendor latency figures are not always directly comparable.
• –Accuracy drops to 85.7% on codes and addresses, where support and commerce agents cannot afford silent mistakes.
• –The model costs $15 per million characters but has no stop message, requiring connection closure for interruption handling.
• –Cekura tests 104 agent sentences across dozens of models, but its accuracy score remains an ASR-based estimate rather than human judgment.