Voice AI BenchmarksBy

MODEL DIRECTORY

Compare voice AI models

Independent benchmarks for 57 speech-to-text, text-to-speech and speech-to-speech models, re-measured every day.

Fastest final transcript (TTFS)#1 / 24
64ms
STT RT v5via Soniox
Fastest first audio (TTFA)#1 / 30
116ms
Palabra TTS v1via Palabra
Fastest voice-to-voice reply (V2V)#1 / 2
1306ms
GPT Realtime 2via OpenAI

57 models

Speech-to-Text: latency vs quality

Speed against accuracy for every measured STT model — the best sit toward the bottom left.

TTFS vs WEREach point is one measured STT model · 30-day averagesTime to Final Segment against Word Error Rate for every measured STT model.
STT RT v5 — 64 ms, 5.8%Parakeet TDT 0.6B v3 — 70 ms, 11.0%STT 1 — 83 ms, 4.8%Nova 3 — 99 ms, 6.3%Nova 2 — 101 ms, 7.9%Ink 2 — 108 ms, 4.9%Scribe v2 Realtime — 120 ms, 5.5%Universal 3.5 Pro — 146 ms, 3.2%Default via Azure — 155 ms, 5.3%Whisper Large v3 — 180 ms, 8.2%Velma 2 STT Streaming — 191 ms, 6.4%Grok STT — 199 ms, 4.8%Pulse — 205 ms, 5.3%Default via Speechmatics — 223 ms, 5.5%Default via Gradium — 249 ms, 10.0%resonant-1 — 288 ms, 3.4%Enhanced — 341 ms, 4.3%Voxtral Mini Transcribe Realtime 2602 — 356 ms, 6.2%GPT Realtime Whisper — 559 ms, 5.2%GPT-4o mini Transcribe — 623 ms, 5.1%Solaria 1 — 680 ms, 7.1%GPT-4o Transcribe — 742 ms, 4.7%Chirp 2 — 811 ms, 5.0%Chirp 3 — 813 ms, 4.1%

Text-to-Speech: latency vs quality

Speed against intelligibility for every measured TTS model — the best sit toward the bottom left.

TTFA vs WEREach point is one measured TTS model · 30-day averagesTime to First Audio against Word Error Rate for every measured TTS model.
Palabra TTS v1 — 116 ms, 5.9%vui — 124 ms, 7.9%TTS Flash 2 — 128 ms, 4.9%TTS 2 — 176 ms, 4.8%Blizzard — 235 ms, 7.4%Neural — 236 ms, 4.3%Mist v3 — 256 ms, 6.3%TTS Rt v2 — 262 ms, 3.9%Sonic 3.5 — 274 ms, 6.1%TTS RT v1 — 275 ms, 3.7%Dragon HD Latest — 310 ms, 5.1%Coda — 313 ms, 5.2%Aura 2 — 328 ms, 5.3%S2.1 Pro — 374 ms, 4.6%Default — 384 ms, 4.6%Speech 2.8 Turbo — 411 ms, 4.8%Eleven v3 Conversational — 412 ms, 4.4%Grok TTS — 420 ms, 4.7%S1 — 434 ms, 4.9%Flash v2.5 — 455 ms, 6.7%Speech 2.8 HD — 460 ms, 4.7%Sonic 3.6 — 466 ms, 5.5%Simba 3.2 — 484 ms, 4.7%Chirp 3 HD — 512 ms, 5.2%Simba 3.0 — 526 ms, 5.3%Falcon 2 — 549 ms, 5.2%Lightning v3.1 Pro — 586 ms, 4.4%Qwen3 TTS Flash Realtime — 645 ms, 8.8%S2.1 Pro Free — 806 ms, 4.7%GPT-4o mini TTS — 1075 ms, 4.8%

Speech-to-Speech: latency vs quality

Speed against instruction adherence for every measured S2S model — the best sit toward the top left.

V2V vs InstructionEach point is one measured S2S model · 30-day averagesVoice-to-Voice Latency against Instruction Adherence for every measured S2S model.
GPT Realtime 2 — 1306 ms, 71%Gemini 3.1 Flash Live (Preview) — 1380 ms, 76%

About this directory

A placement always names its metric — #2 on TTFS is a different claim than #2 on WER — and is judged across every host serving the model. Models without enough qualifying samples in the last 30 days show no number at all.

Open-weight models can post different latency on different hosts because serving infrastructure and configuration differ, so each hosted endpoint is measured separately. The provider directory covers which organizations create models, host them, or do both.

Same datasets, prompts and metric definitions for every model, measured by Coval’s open-source runner. Full methodology on the overview.

Evaluate your own voice agent

Use Coval to test your production configuration, prompts and calls—not only the public benchmark endpoints.

Book a Demo