Text-to-Speech (TTS) benchmark — by Coval
An independent benchmark of text-to-speech providers on Time to First Audio (TTFA) and Word Error Rate (WER), measured under production-realistic conditions by Coval. Fastest is eleven_flash_v2_5 (212 ms median TTFA), lowest word error rate is octave-2 (3.8%). Mirrored here with attribution; last 7d.
Run by Coval, a voice-AI evaluation platform
This benchmark is produced by Coval, a voice-AI evaluation platform used by teams like Perplexity, Zoom, ServiceNow, and GEICO to test production voice agents. Coval builds testing and evaluation infrastructure for voice AI — it does not sell TTSmodels of its own — so it measures every provider as a neutral third party rather than benchmarking its own product. Every number here is Coval's, measured on a pinned dataset under production-realistic conditions; the methodology, runner code, and inputs are open-source (Apache-2.0) and reproducible by re-running the suite. Openbenchmarks mirrors these results daily with attribution.
Text-to-Speech models ranked by Time to First Audio
Every model runs the same pinned dataset under the same conditions. TTFA (median) is the headline latency; WER is accuracy (lower is better on both). Ranked fastest-first, last 7d:
| # | Model | TTFA median | TTFA p95 | WER | Samples |
|---|---|---|---|---|---|
| 1 | eleven_flash_v2_5 Elevenlabs | 212 ms | 2,849 ms | 5.5% | 3,360 |
| 2 | inworld-tts-1.5-mini Inworld | 228 ms | 480 ms | 6.0% | 3,360 |
| 3 | neural Azure | 235 ms | 295 ms | 4.3% | 990 |
| 4 | s2.1-pro Fishaudio | 239 ms | 447 ms | 5.7% | 990 |
| 5 | inworld-tts-1.5-max Inworld | 250 ms | 429 ms | 5.7% | 3,360 |
| 6 | mistv3 Rime | 259 ms | 291 ms | 6.3% | 3,359 |
| 7 | sonic-3.5 Cartesia | 276 ms | 351 ms | 5.7% | 3,357 |
| 8 | dragon-hd-latest Azure | 286 ms | 381 ms | 4.9% | 990 |
| 9 | tts-rt-v1 Soniox | 293 ms | 330 ms | 5.0% | 3,357 |
| 10 | coda Rime | 301 ms | 369 ms | 7.3% | 3,335 |
| 11 | aura-2-thalia-en Deepgram | 329 ms | 538 ms | 7.7% | 3,354 |
| 12 | default Gradium | 331 ms | 407 ms | 3.9% | 3,360 |
| 13 | arcana Rime | 334 ms | 380 ms | 7.6% | 3,360 |
| 14 | grok-tts Xai | 366 ms | 464 ms | 4.0% | 3,354 |
| 15 | s1 Fishaudio | 440 ms | 621 ms | 5.5% | 990 |
| 16 | chirp-3-hd | 452 ms | 815 ms | 5.8% | 990 |
| 17 | sonic-3 Cartesia | 456 ms | 808 ms | 6.2% | 3,356 |
| 18 | eleven_v3 Elevenlabs | 589 ms | 3,198 ms | 5.4% | 1,009 |
| 19 | octave-2 Hume | 592 ms | 805 ms | 3.8% | 3,344 |
| 20 | octave-tts Hume | 673 ms | 945 ms | 4.2% | 3,346 |
| 21 | qwen3-tts-flash-realtime Alibaba | 689 ms | 756 ms | 9.1% | 990 |
| 22 | gpt-4o-mini-tts Openai | 812 ms | 1,604 ms | 4.8% | 3,074 |
| 23 | lightning_v3.1_pro Smallest | 939 ms | 1,132 ms | 7.5% | 3,358 |
| 24 | eleven_multilingual_v2 Elevenlabs | 1,267 ms | 1,637 ms | 4.4% | 3,081 |
Numbers are point-in-time against Coval's pinned dataset and refresh continuously — they don't generalize indefinitely. Full charts, distributions, and windows (24h / 7d / 30d) at benchmarks.coval.ai →