Does this site run its own tests?

No. Every number is copied from one of three public benchmarks, Coval, Speko and Artificial Analysis, read on the date shown, with a link to the source.

Why is there no overall score?

The three benchmarks test different models under different conditions. Adding or averaging their ranks would hide that, so models are grouped by speed first and every other measure is shown side by side.

What happens when a benchmark does not measure a model?

Nothing is estimated. The model page lists the missing measure under “Not collected”, with the reason. A model Coval does not measure has no speed tier.

How often is the data read again?

The three sources are captured again by script; each capture is kept with its date and checksum, and every page shows the date of the reading it uses (currently October 22, 2026).

Methodology

Sources and methodology: three independent benchmarks, speed first, no combined score

All three benchmarks read on October 22, 2026. TTS for Voice Agents compares 119 text-to-speech (TTS) models for real-time voice agents. It runs no tests of its own: this page says where every number comes from, how it was read, and the rules applied to it.

The three benchmarks

Source	What it measures	How	Models	Used for
Coval	Time to first audio, word error rate	Calls each hosted API every 30 minutes from US East; publishes percentiles over 30 days	32	Speed (median and P95), speed tiers, feature tags (multilingual), some prices
Speko	Naturalness, time to first audio, pronunciation of numbers and dates, cost, voice drift	Blind A/B votes on phone-agent lines; latency sweeps from US East	45	Naturalness, number pronunciation, the speed and naturalness frontier
Artificial Analysis	Listener preference (Elo), list price	Blind listening arena with a 95% interval per model; votes can be filtered by category	98	Listener Elo, the second frontier, the customer service table, prices

Rules applied

Rule	Definition	Reason
Speed first	Models are grouped by speed before anything else is compared.	The site is written for real-time voice agents. This is an editorial choice, stated at the top of the speed page.
Two percentiles	Median (P50) and P95, both from Coval's 30-day window.	The median is the typical wait, P95 the slow turn. Showing one alone would favour whichever model leads on it.
Tie rule, latency	Two latencies within 10 ms are a tie.	The order among the leaders changes from week to week on gaps of a few milliseconds.
Tie rule, ratings	Two Elo scores are a tie when their published 95% intervals overlap.	The source's own measure of uncertainty.
Speed tiers	Headroom: P95 at or under 101 ms; Fast enough: P95 from 138 ms to 263 ms; Visible delay: P95 of 320 ms or more.	The breaks follow gaps in the data: no model sits between two tiers. They are recomputed at each reading.
Frontier	A model is on the frontier when no other model is both faster and better rated.	Compares two measures without weights or score cut-offs.
No thresholds	Each comparison includes every model the sources involved both list.	A cut-off can be placed to include or exclude a given model.
No combined score	Ranks and scores from different benchmarks are never added or averaged.	The three sources test different model sets under different conditions.
Accuracy measure	Speko's number and date test, not Coval's word error rate.	On Coval's 30-day table, 25 of 32 models fall between 1.8% and 2.8%: too close to rank.
Prices	List price per 1M characters as each source prints it; a range when sources or plans differ.	Plans differ; Gradium's plans are read on gradium.ai/pricing and converted at one credit per character.

Headroom (4 models): Headroom to spend elsewhere in the turn.
Fast enough (9 models): Fast enough for most agents.
Visible delay (19 models): The voice becomes a visible part of the delay.

How each source was read

Source	Settings	What was taken
Coval dashboard and API	Window 30 days; percentiles P50 and P95; no other filter	All 32 models measured
Coval model registry and pages	None	Feature tags (multilingual, voice cloning, emotion control), licensing, creator and host
Coval pricing	Text-to-speech table	List price per 1M characters and the date it applies from
Speko TTS board	Region US East; all models shown	Naturalness, time to first audio (board chart), pronunciation score, cost; 41 models have both latency and naturalness
Artificial Analysis, Provider Voice	Accent All; Category All	Elo, 95% interval, rank and price for the 98 models rated
Artificial Analysis, Provider Voice	Category Customer Service	Rank, Elo and interval for the 93 models rated in that category
gradium.ai	None	Plans and credits, converted to a price per 1M characters

Each source is captured by script on the date of the reading; the raw files are kept with a checksum, so every number can be traced back to the page it was read from.

Model matching

Models are matched across sources by vendor and model name. When a benchmark measures a model through another company's API, the host is named on the model page and the latency is the host's (5 models).

Some matches are inferred rather than stated by any source, and the model page says so. Gradium TTS is matched across the three sources from the date in each name: Coval's “Default” model (gradium-tts-beta-202609), Speko's gradium:default and Artificial Analysis's “Gradium TTS (Sep 2026)”.

Known limits of the data

All latency is measured from US East, in English. No public source measures from Europe.
The benchmarks update daily. Figures are a reading on the date shown, and the order among the fastest models changes within a week.
Coverage differs: a model can be measured by one source and absent from the others. A missing value is listed under “Not collected” on the model page, with the reason; it is never estimated.
Speko marks some naturalness scores as provisional and can report two latency figures for one model (the board chart and the model page sweep); both are shown.
Coval no longer publishes the P99 in its 30-day aggregates; the P95 is the slow-turn measure used.

Questions and answers

Does this site run its own tests?: No. Every number is copied from one of three public benchmarks, Coval, Speko and Artificial Analysis, read on the date shown, with a link to the source.
Why is there no overall score?: The three benchmarks test different models under different conditions. Adding or averaging their ranks would hide that, so models are grouped by speed first and every other measure is shown side by side.
What happens when a benchmark does not measure a model?: Nothing is estimated. The model page lists the missing measure under “Not collected”, with the reason. A model Coval does not measure has no speed tier.
How often is the data read again?: The three sources are captured again by script; each capture is kept with its date and checksum, and every page shows the date of the reading it uses (currently October 22, 2026).

Change log

October 22, 2026: methodology version 2. Speed tiers, tie rules and frontiers replace the score out of 100; new reading of the three sources.
October 5, 2026: first version.

Published by Rank-ly. Corrections: contact. Data from Coval, Speko and Artificial Analysis belongs to its publishers.

Get benchmark updates

An occasional email when the sources are read again.

Benchmark updates only. Unsubscribe at any time. See our privacy policy.