The TTS benchmark for real-time voice agents Real-time text-to-speech (TTS) APIs, grouped by measured speed, then compared on voice quality, accuracy and price.

Frequently asked questions

Choosing a TTS API for a voice agent

What is the best TTS for voice agents in 2026?

This benchmark does not crown one: it groups models by measured speed first, then compares voice quality, accuracy and price without adding them up. On Coval's 30-day window read on October 22, 2026, 4 models start speaking in 101 ms or less, even on slow turns (P95): Fluxions vui, Alibaba Qwen3 TTS Fast, Inworld Realtime TTS-2 Flash and Gradium TTS. On slow turns (P95), Fluxions vui is ahead on its own; Alibaba Qwen3 TTS Fast, Inworld Realtime TTS-2 Flash and Gradium TTS are a tie. On typical turns (median), Fluxions vui is ahead on its own; Gradium TTS, Inworld Realtime TTS-2 Flash and Alibaba Qwen3 TTS Fast are a tie. Gaps under 10 ms are treated as ties. They differ on voice ratings, numbers and price: see them side by side on the real-time page.

How is this text-to-speech benchmark built?

It reads three independent benchmarks and runs no tests of its own: Coval (time to first audio, median and P95, over 30 days), Speko (naturalness, number and date pronunciation, latency, cost) and Artificial Analysis (blind listener preference with a 95% interval, and price). Models are grouped by speed first; latencies within 10 ms and Elo scores whose intervals overlap are ties; scores from different benchmarks are never added or averaged.

What is time-to-first-audio (TTFA), and why the median and P95?

Time-to-first-audio is the wait between sending text to a TTS API and hearing the first syllable. For a voice agent it decides whether a reply feels instant. The median (P50) is the typical turn; the P95 is the slowest one turn in twenty, the one a caller notices. Showing one alone would favour whichever model leads on it.

Which TTS API has the lowest latency?

On slow turns (P95), Fluxions vui (83 ms) is ahead on its own; on typical turns (median), Fluxions vui (41 ms) is ahead on its own. Gaps under 10 ms are ties.

Which TTS models are open source?

22 models in the benchmark publish their weights (open weights), according to Artificial Analysis, the Coval registry or the creator's own release. They are listed on the open-weights page, with the same measurements as the proprietary APIs.