Its leaderboard reports Inworld TTS-2 Flash at 45 ms streaming latency, 93.2% accuracy, and $15 per million characters; Gradium TTS is reported to begin speaking before the final word arrives on tested streaming inputs, with 79.5% accuracy. The page notes that accuracy and naturalness are estimates and that eight voices unable to take streamed text are not shown in the streaming chart. The benchmark gives builders a comparative starting point, but its measures and test conditions should not be treated as a universal ranking. Source: https://benchmarks.cekura.ai/text-to-speech
26 Best Text-to-Speech Models for Voice Agents 2026 (Tested)
In one minute
Cekura has published a text-to-speech benchmark for voice agents, presenting results for 26 models and separating streaming latency, accuracy, naturalness, price, and stopping behavior.
What changed: Cekura adds a 26-model benchmark focused on voice-agent TTS, comparing streaming latency, accuracy, naturalness, pricing and stop behavior; the prior Open TTS leaderboard covers multilingual TTS and voice cloning.