Does this site run its own TTS tests?

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

Why is there no overall score for TTS models?

The four benchmarks test different model sets under different conditions. Adding or averaging their results would hide that, so models are ordered by speed and every other measure is shown side by side.

How are ties handled in the ranking?

Two P95 latencies within 10 ms share a rank. Tied models are displayed by median latency, a rule applied to every tie. The choice of this display order is the publisher's; the rank and the measurements are unchanged by it.

Why are some TTS models not ranked?

A model is ranked when Coval tags it multilingual and measures its latency: 23 models. The other 73 keep a page with every measure we hold, and the missing ones are listed with the reason.

How do I ask for a correction?

Use the contact page. A figure is corrected against its public source, and a vendor can ask for a right of reply on its own model page.

Data read October 27, 2026

Methodology: how the multilingual TTS API ranking is built

This site runs no tests of its own. It copies figures from four public benchmarks, orders 23 multilingual real-time TTS APIs by speed, and shows voice quality, accuracy and price beside that order without ever combining them. This page says where each number comes from and which rules were applied.

Who publishes it

The site is published by Archipel Marketing (SAS, SIREN 984715995). Corrections and rights of reply go through the contact page.

The four sources

Source	What it measures	How	Models used	Used here for
Coval	Time to first audio and word error rate	Calls each hosted API every 30 minutes from the US East and publishes percentiles over 30 days	31	Median and P95 latency, speed tiers, the multilingual tag, some prices
Speko	Naturalness, numbers and dates, cost	Blind A/B votes on phone-agent lines, latency runs from the US East	43	Naturalness and numbers-and-dates pronunciation
Artificial Analysis	Listener preference (Elo), hard-passage pronunciation, list price	Blind listening arena with a 95% interval per model; human reviewers judge hard passages	73	Listener Elo, customer-service Elo, pronunciation, prices, languages tested, use-case categories
Cekura	First-audio time with and without streamed text, reading accuracy, estimated naturalness, price	Each model reads the same phone-agent sentences from one cloud region; two speech-to-text systems score the reading	18	Reading accuracy, streamed-text input, estimated naturalness, prices

All readings are from October 27, 2026. The benchmarks update daily, so the order among the fastest models can move within a week. Each model page lists the exact URLs and the date each was read.

Which models are ranked

The database holds 96 models from 35 providers. 23 are ranked: the models Coval tags multilingual and for which Coval measures time to first audio. The other 73 stay in the database, with a page and every measure we hold, but have no rank. A model that Coval does not measure has no speed tier.

The rules

Speed first. Models are grouped into speed tiers from the Coval P95 over 30 days, then ordered by P95 inside a tier. Headroom: P95 75.0 to 104.3 ms. Fast enough: 136.6 to 267.0 ms. Visible delay: 319.0 to 1,137.9 ms (ranges recounted from the data).
Tier boundaries. Each break sits in the widest gap between two consecutive P95 values, so it does not depend on any one model.
Ties. Two P95 values within 10 ms share a rank, because the order among the leaders changes on gaps of a few milliseconds.
Display order inside a tie. Tied models are listed by median latency, lowest first. This display rule is the publisher's choice, and it is applied to every tie. At rank 1 it lists Gradium TTS, then Inworld Realtime TTS-2 Flash, then Alibaba Qwen3 TTS Fast. Listed by P95 instead, or alphabetically, the order of those 3 models would be different; their shared rank and their measurements would not.
No combined score. Ranks and scores from different benchmarks are never added or averaged.
Voice quality. Listener Elo, Speko naturalness and Cekura naturalness stay in separate columns. Two Elo values whose 95% intervals overlap are tied.
Accuracy. Cekura reading accuracy, Artificial Analysis hard-passage pronunciation and the Speko numbers-and-dates score use different texts and judges, so they are shown side by side.
Prices. List price per million characters as each benchmark prints it, with a range when they differ. No price is read from a vendor site.
Streaming only. A model is listed when a benchmark times its first audio or when the vendor's documentation describes a streaming API. A voice agent cannot wait for a whole sentence to be synthesized.

What is deduced rather than declared

The Best for column on the ranking is computed: it names the measures on which a model leads among the ranked models. It is not a vendor claim.
Model summaries are generated from the measures shown on the page.
A language is listed for a model when Artificial Analysis or Speko tested it in that language. This is evidence of support, not a quality measure.

Known limits

All latency is measured from the US East, in English. No public source measures from Europe or Asia, and none publishes speed, naturalness or accuracy per language.
"Multilingual" is Coval's tag. It says a vendor supports several languages, not that quality holds in each.
A model measured by one source and absent from the others has n/a in the other columns. Nothing is estimated; the model page lists each missing measure with its reason.
Cekura's naturalness is a listener model's estimate and its accuracy is scored by transcribers, not people.

Frequently asked questions

Does this site run its own TTS tests?: No. Every number is copied from one of four public benchmarks, Coval, Speko, Artificial Analysis and Cekura, read on the date shown, with a link to the source on each model page.
Why is there no overall score for TTS models?: The four benchmarks test different model sets under different conditions. Adding or averaging their results would hide that, so models are ordered by speed and every other measure is shown side by side.
How are ties handled in the ranking?: Two P95 latencies within 10 ms share a rank. Tied models are displayed by median latency, a rule applied to every tie. The choice of this display order is the publisher's; the rank and the measurements are unchanged by it.
Why are some TTS models not ranked?: A model is ranked when Coval tags it multilingual and measures its latency: 23 models. The other 73 keep a page with every measure we hold, and the missing ones are listed with the reason.
How do I ask for a correction?: Use the contact page. A figure is corrected against its public source, and a vendor can ask for a right of reply on its own model page.