Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
TTS reading and pronunciation accuracy in 2026: numbers, dates and hard passages
12 of the 23 ranked models have a Cekura reading-accuracy result; Inworld Realtime TTS-2, Inworld Realtime TTS-2 Flash, and Fish Audio S2.1 Pro are highest, starting with the best. Artificial Analysis and Speko test pronunciation on other texts and are shown beside it.
Accuracy matters for phone agents that read amounts, dates and reference codes aloud. The tests use different texts and judges, so no combined accuracy is computed. Coval's word error rate is shown for reference only.
How the rows are ordered. Ordered by Cekura reading accuracy, highest first; models without a Cekura result come last, in ranking order.
| Model | Reading accuracy (Cekura) Share of test sentences read correctly | Numbers (Cekura) | Codes (Cekura) | Hard passages (AA) Artificial Analysis Pronunciation Robustness, share judged correct | Numbers and dates (Speko) Speko pronunciation score, as published | Word error rate (Coval) | Rank |
|---|---|---|---|---|---|---|---|
 Inworld Realtime TTS-2 | 93.4% | 93% | 86% | 80.7% | 0.61 | 1.7% | 6 |
 Inworld Realtime TTS-2 Flash | 93.2% | 93% | 86% | 77.8% | 0.58 | 1.7% | 1 = |
 Fish Audio S2.1 Pro | 92.0% | 90% | 86% | 88.1% | n/a | 1.6% | 19 |
 SpaceXAI Grok TTS | 87.4% | 82% | 86% | 87.6% | 0.45 | 1.7% | 18 |
 Soniox TTS Real-Time v2 | 83.6% | 78% | 86% | 78.8% | 0.49 | 2.6% | 10 = |
 ElevenLabs Eleven v4 Turbo | 79.9% | 71% | 82% | 90.1% | 0.90 | 1.7% | 8 |
 Gradium TTS | 79.5% | 77% | 77% | n/a | 0.27 | 1.9% | 1 = |
 Rime Mist v3 | 77.2% | 71% | 77% | n/a | n/a | 2.1% | 12 = |
 Rime Coda | 76.7% | 70% | 79% | n/a | n/a | 11.2% | 16 = |
 Cartesia Sonic 3.6 | 74.9% | 71% | 75% | 74.5% | 0.21 | 1.5% | 14 |
 Cartesia Sonic 3.5 | 73.1% | 69% | 75% | 68.8% | 0.18 | 1.6% | 15 |
 ElevenLabs Flash v2.5 | 64.4% | 55% | 55% | n/a | n/a | 1.9% | 9 |
 Alibaba Qwen3 TTS Fast | n/a | n/a | n/a | n/a | 0.30 | 1.9% | 1 = |
 Alibaba Qwen3 TTS 1.7b | n/a | n/a | n/a | n/a | n/a | 1.9% | 4 |
 SpeechifyAI Simba 3.0 | n/a | n/a | n/a | n/a | n/a | 1.6% | 5 |
 Palabra TTS v1 | n/a | n/a | n/a | n/a | 0.16 | 2.8% | 7 |
 Soniox TTS RT v1 | n/a | n/a | n/a | n/a | 0.35 | 2.6% | 10 = |
 Murf AI Falcon 2 | n/a | n/a | n/a | n/a | n/a | 1.7% | 12 = |
 Deepdub PhantomZ 4.0 | n/a | n/a | n/a | n/a | n/a | 2.3% | 16 = |
 Smallest.ai Lightning V3.1 Pro (Jul 2026) | n/a | n/a | n/a | 72.1% | n/a | 1.6% | 20 |
 Alibaba Qwen3 TTS Flash Realtime | n/a | n/a | n/a | n/a | n/a | 3.1% | 21 |
 Google Chirp 3: HD | n/a | n/a | n/a | n/a | n/a | 1.6% | 22 |
 Fish Audio S2.1 Pro Free | n/a | n/a | n/a | n/a | n/a | 1.6% | 23 |
Frequently asked questions
- Which TTS API reads numbers and codes most accurately?
- On Cekura reading accuracy, Inworld Realtime TTS-2 is highest among ranked models (93.4%). Check the numbers and codes columns, which can order models differently from the overall figure.
- How is TTS accuracy measured?
- Cekura has each model read the same phone-agent sentences and two speech-to-text systems score the reading. Artificial Analysis has human reviewers judge hard passages. Speko scores numbers and dates.
- Why do some models have no accuracy result?
- No benchmark tested them. The model page lists each missing measure under not collected, with the reason; nothing is estimated.
- Is accuracy measured per language?
- No. The accuracy tests are not published per language.