Methodology · read on October 27, 2026
Sources and methodology: four public benchmarks, speed first, no combined score
This site ranks 96 text-to-speech (TTS) models for the US. It runs no test of its own: this page says where every number comes from, how it was read and which rules were applied to it. Figures are those of October 27, 2026. The site is published by Archipel Marketing (SAS, SIREN 984715995).
The four benchmarks
| Source | What it measures | How | Models | Used for |
|---|---|---|---|---|
| Coval | Time to first audio and word error rate | Calls each hosted API every 30 minutes from US East and publishes percentiles over 30 days. | 31 | Median and P95 speed, speed tiers, feature tags, some prices. |
| Speko | Naturalness, time to first audio, pronunciation of numbers and dates, cost | Blind A/B votes on phone-agent lines; latency sweeps from US East. | 45 | Naturalness and number pronunciation. |
| Artificial Analysis | Listener preference (Elo), pronunciation of hard passages, list price | Blind listening arena with a 95% interval per model; human reviewers judge hard passages. | 73 | Listener Elo, the four use case categories, pronunciation, prices. |
| Cekura | Time to first audio with and without streamed text, reading accuracy, estimated naturalness, price | Every model reads the same phone-agent sentences from one cloud region; two speech-to-text transcribers score the reading. | 26 | Reading accuracy, latency, streaming status, prices. |
Rules applied
| Rule | Definition | Reason |
|---|---|---|
| Speed first | Models are grouped by speed before anything else is compared. | This is an editorial choice: the site is written for real-time voice products. |
| Two percentiles | Median (P50) and P95, both from Coval's 30-day window. | The median is the typical wait, P95 the slow turn. One alone would favor whichever model leads on it. |
| Tie rule, latency | Two latencies within 10 ms are a tie, ordered by median time to first audio. | 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. | It is the source's own measure of uncertainty. |
| Speed tiers | Headroom: P95 under 120 ms. Fast enough: 120 to 293 ms. Visible delay: above 293 ms. | Each break is the middle of the widest gap between consecutive P95 values, searched between 80 and 160 ms and between 200 and 400 ms, so no model decides it. Recomputed at each reading. |
| No thresholds | A comparison includes every model the sources 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 four sources test different model sets under different conditions. |
| Streaming only | A model is listed when a benchmark times its first audio, or when its vendor's documentation describes a streaming API. | A voice agent cannot wait for a whole sentence to be synthesized. |
| Accuracy side by side | Speko's numbers and dates, Artificial Analysis's pronunciation of hard passages and Cekura's reading accuracy are shown separately. Coval's word error rate is not used to compare. | The tests use different texts and judges. On Coval's table, most models fall within a point of each other. |
| Prices | List price per million characters as each benchmark prints it; a range when they differ. | No price is read from a vendor's own site: every model is priced from the same benchmarks. |
Speed tiers
- Headroom (4 models): there is headroom to spend elsewhere in the turn.
- Fast enough (9 models): fast enough for most agents.
- Visible delay (18 models): the voice becomes a visible part of the delay.
Only 31 of the 96 models are timed by Coval, so only they have a tier. The tier is computed here from Coval's P95; the vendors do not declare it.
Use cases
Four use cases are the listening categories of Artificial Analysis's arena: customer service, assistants, entertainment and knowledge sharing. A model is listed on a use case page when Artificial Analysis rates it for that category. The fifth, voice agents, is computed here: it holds the 13 models whose P95 stays under 293 ms, that is the Headroom and Fast enough tiers. Models are ordered on each page by the same rule: speed tier, P95, then listener Elo for models Coval does not time (customer service 70, voice agents 13, assistants 69, entertainment 69, knowledge sharing 73 models).
Models with neither a Coval timing nor an Artificial Analysis rating (10 of them) have no rank. They are listed after the ranked models, in alphabetical order, and the page says so.
Comparison pages
Each comparison page sets one reference model, Gradium TTS, against another listed model, so there are 95 of them. The reference model is the editor's choice, not a result of the ranking, and it is the same on every comparison page. Every sentence on a comparison page is computed from the figures in its table; when a measure is missing for one of the two models, the page says it does not compare them on it.
What is stated and what is deduced
- Stated by a benchmark or a vendor: every measurement, price, language and streaming status, with its source and reading date on the model page.
- Computed by this site: speed tiers, ranks, the voice agents list, ties and the comparison sentences.
- Inferred: the match between a model and benchmark entries that spell its name differently. They are listed below and on the model page concerned.
How each source was read
| Source | Settings | What was taken |
|---|---|---|
| Coval dashboard and API | Window 30 days; percentiles P50 and P95; no other filter | Every model measured (31) |
| Coval model registry and pricing | Text-to-speech table | Feature tags, creator and host, list price and its start date |
| Speko TTS board | Region US East; all models shown | Naturalness, latency, pronunciation score, cost |
| Artificial Analysis, provider voice | Accent all; category all, then each of the four categories | Elo, 95% interval, rank and price; category Elo for customer service |
| Artificial Analysis, pronunciation | All categories | Share of hard passages judged correct and its 95% interval |
| Cekura TTS benchmark | Default view (streamed text) and the per-category accuracy table | Median time to first audio, reading accuracy, estimated naturalness, price |
| Vendor documentation | Streaming API pages only | Streaming status for the 46 models no benchmark times; never used for a measure or a price |
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. Sources do not always spell a model's name the same way: when a name differs (an API identifier, a shortened name, a date instead of a version), the match is inferred and listed here. 31 models have at least one inferred match.
| Model page | Provider | Name as the source writes it |
|---|---|---|
| Alibaba Qwen3 TTS | Alibaba (Qwen) | Cekura: “Qwen3-TTS” |
| Alibaba Qwen3 TTS Fast | Alibaba (Qwen) | Speko: “Nari Qwen3-TTS Fast” |
| Alibaba Qwen3 TTS Flash | Alibaba (Qwen) | Speko: “Qwen qwen3-tts-flash” |
| Bland Speech v3 | Bland AI | Speko: “Bland speech” |
| Deepgram Flux TTS | Deepgram | Cekura: “Deepgram Flux”; Speko: “Deepgram flux” |
| ElevenLabs Eleven v3 | ElevenLabs | Cekura: “ElevenLabs v3” |
| ElevenLabs Eleven v3 Conversational | ElevenLabs | Speko: “ElevenLabs eleven_v3_conversational” |
| ElevenLabs Eleven v4 | ElevenLabs | Cekura: “ElevenLabs v4” |
| ElevenLabs Eleven v4 Turbo | ElevenLabs | Cekura: “ElevenLabs v4 Turbo” |
| ElevenLabs Flash v2.5 | ElevenLabs | Speko: “ElevenLabs eleven_flash_v2_5” |
| ElevenLabs Multilingual v2 | ElevenLabs | Speko: “ElevenLabs eleven_multilingual_v2” |
| ElevenLabs Turbo v2.5 | ElevenLabs | Speko: “ElevenLabs eleven_turbo_v2_5” |
| Fish Audio Drama 3 Preview | Fish Audio | Cekura: “Fish Audio Drama 3”; Coval: “Fish Audio drama3-preview” |
| Fish Audio S2.1 Pro | Fish Audio | Artificial Analysis (pronunciation): “Fish Audio S2.1 Pro (Oct 2026)”; Artificial Analysis: “Fish Audio S2.1 Pro (Oct 2026)” |
| Google Gemini 3.1 Flash TTS | Cekura: “Gemini 3.1 Flash”; Speko: “Gemini 3.1 flash tts preview” | |
| Google Gemini 3.8 Flash TTS | Cekura: “Gemini 3.8 Flash”; Speko: “Gemini 3.8 flash tts” | |
| Google Gemini 3.8 Flash-Lite TTS | Cekura: “Gemini 3.8 Flash-Lite”; Speko: “Gemini 3.8 flash lite tts” | |
| Gradium TTS | Gradium | Artificial Analysis: “Gradium TTS (Sep 2026)”; Cekura: “Gradium TTS”; Coval: “Gradium Default”; Speko: “Gradium TTS” |
| Hume AI Octave 2 | Hume AI | Cekura: “Hume Octave 2”; Speko: “Hume octave-2” |
| Hume AI Octave TTS | Hume AI | Speko: “Hume octave-1” |
| Inworld Realtime TTS-2 | Inworld | Cekura: “Inworld TTS-2”; Coval: “Inworld AI TTS 2”; Speko: “Inworld tts 2” |
| Inworld Realtime TTS-2 Flash | Inworld | Cekura: “Inworld TTS-2 Flash”; Coval: “Inworld AI TTS Flash 2”; Speko: “Inworld tts 2 flash” |
| Murf AI Falcon 2 | Murf AI | Coval: “Murf Falcon 2” |
| OpenAI GPT-4o mini TTS | OpenAI | Cekura: “GPT-4o mini TTS” |
| Palabra TTS v1 | Palabra | Coval: “Palabra TTS” |
| Rime Mist v3 | Rime | Speko: “Rime mistv3” |
| Smallest.ai Lightning v3.1 | Smallest.ai | Cekura: “Smallest Lightning v3.1”; Speko: “Smallest lightning_v3.1” |
| Smallest.ai Lightning V3.1 Pro (Jul 2026) | Smallest.ai | Coval: “Smallest Lightning v3.1 Pro”; Speko: “Smallest lightning_v3.1_pro” |
| Soniox TTS Real-Time v2 | Soniox | Cekura: “Soniox TTS RT v2”; Coval: “Soniox TTS Rt v2”; Speko: “Soniox tts-rt-v2” |
| SpaceXAI Grok TTS | xAI | Cekura: “Grok TTS”; Coval: “SpaceXAI Grok TTS”; Speko: “xAI Grok grok-tts” |
| SpeechifyAI Simba 3.2 | Speechify | Speko: “Speechify simba-3.2” |
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 can change within a week.
- Coverage differs: a model can be measured by one source and absent from the others. A missing value is listed as not collected on the model page, with the reason, and is never estimated.
- Speko marks some naturalness scores as provisional and can report two latency figures for one model; both are shown.
- 46 models stream according to their vendor's documentation but are not timed by any benchmark: they have no speed tier.
- Cekura's naturalness is a listener model's estimate and its accuracy is scored by transcribers, not people.
Data from Coval, Speko, Artificial Analysis and Cekura belongs to its publishers. To report an error or answer a listing, use the contact page.
Questions and answers
- Does this site run its own tests?
- No. Every number is copied from one of four public benchmarks (Coval, Speko, Artificial Analysis and Cekura), read on October 27, 2026, with a link to the source.
- Why is there no overall score?
- The four 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.
- Why are some TTS models not listed?
- Only models that stream audio are listed: a voice agent cannot wait for a whole sentence to be synthesized. A model is kept when a benchmark times its first audio or its vendor's documentation describes a streaming API.
- 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 time has no speed tier.
- Who publishes this site?
- Archipel Marketing (SAS, SIREN 984715995). Corrections and replies from a listed company go through the contact page.