BENCHMARKSTT24 MODELS RANKEDLAST 30 DAYS

Time to First Token (TTFT)

Name: Time to First Token voice AI benchmark results
Creator: Coval

Time to First Token (TTFT) is how quickly a speech-to-text model starts streaming partial transcripts after audio is sent.

Current STT leader#1 / 24: 1030ms; Universal 3.5 Provia AssemblyAI

How it is calculated

TTFT = first partial-transcript timestamp − start of the measured audio request.

How to read it

Unit: ms
Better: Lower
Period: Rolling 30 days
Cadence: Re-measured daily

Speech-to-Text models on Time to First Token

Full STT dashboard

Time to First Token — every measured STT modelMilliseconds · lower is better · 30-day average

#1Universal 3.5 Pro1030 ms
#2Flux1089 ms
#3Flux Multilingual1175 ms
#4Parakeet TDT 0.6B v31210 ms
#5Whisper Large v31241 ms
#6Nova 31434 ms
#7Nova 21436 ms
#8STT 11468 ms
#9Defaultvia Speechmatics1473 ms
#10Universal Streaming1510 ms
#11STT RT v51533 ms
#12Enhanced1536 ms

Show all 24 models

#13Nemotron 3.5 ASR Streaming1540 ms
#14Velma 2 STT Streaming1576 ms
#15Solaria 11703 ms
#16Defaultvia Azure1792 ms
#17Ink 21812 ms
#18GPT Realtime Whisper1823 ms
#19Voxtral Mini Transcribe Realtime 26021828 ms
#20Defaultvia Gradium1979 ms
#21Pulse2072 ms
#22Scribe v2 Realtime2157 ms
#23Chirp 35999 ms
#24Chirp 26000 ms

median of all models · 1538 ms

STT models over the last 30 days, ranked on Time to First Token.
#	Model	Host	TTFS	TTFT	Samples
1	Universal 3.5 Pro	AssemblyAI	146 ms	1030 ms	29,317
2	Flux	Deepgram	—	1089 ms	29,354
3	Flux Multilingual	Deepgram	—	1175 ms	29,344
4	Parakeet TDT 0.6B v3	Together AI	70 ms	1210 ms	29,187
5	Whisper Large v3	Together AI	180 ms	1241 ms	29,133
6	Nova 3	Deepgram	99 ms	1434 ms	29,335
7	Nova 2	Deepgram	101 ms	1436 ms	29,310
8	STT 1	Inworld AI	83 ms	1468 ms	29,321
9	Default	Speechmatics	223 ms	1473 ms	29,353
10	Universal Streaming	AssemblyAI	—	1510 ms	29,297
11	STT RT v5	Soniox	64 ms	1533 ms	29,363
12	Enhanced	Speechmatics	341 ms	1536 ms	29,356
13	Nemotron 3.5 ASR Streaming	Together AI	—	1540 ms	29,193
14	Velma 2 STT Streaming	Modulate	191 ms	1576 ms	17,919
15	Solaria 1	Gladia	680 ms	1703 ms	26,391
16	Default	Azure	155 ms	1792 ms	6,689
17	Ink 2	Cartesia	108 ms	1812 ms	29,329
18	GPT Realtime Whisper	OpenAI	559 ms	1823 ms	29,327
19	Voxtral Mini Transcribe Realtime 2602	Mistral	356 ms	1828 ms	29,206
20	Default	Gradium	249 ms	1979 ms	28,273
21	Pulse	Smallest	205 ms	2072 ms	29,354
22	Scribe v2 Realtime	ElevenLabs	120 ms	2157 ms	29,245
23	Chirp 3	Google	813 ms	5999 ms	29,360
24	Chirp 2	Google	811 ms	6000 ms	29,278
—	GPT-4o mini Transcribe	OpenAI	623 ms	—	29,359
—	GPT-4o Transcribe	OpenAI	742 ms	—	29,359
—	Grok STT	xAI	199 ms	—	29,363
—	resonant-1	Reson8	288 ms	—	17,091

Under-sampled models are excluded; tied models share a place. Dotted WER values split into substitutions, deletions and insertions on hover or tap.

Tail latency

The distribution behind each average, ranked by median TTFT.

Time to First Token distribution per modelMilliseconds · shared axis · last 30 days

Nova 2p50 1005 ms · p99 3007 ms
Nova 3p50 1010 ms · p99 2990 ms
Universal 3.5 Prop50 1117 ms · p99 2918 ms
Fluxp50 1171 ms · p99 2897 ms
Flux Multilingualp50 1201 ms · p99 2912 ms
Whisper Large v3p50 1320 ms · p99 3480 ms
Parakeet TDT 0.6B v3p50 1337 ms · p99 3042 ms
Nemotron 3.5 ASR Streamingp50 1467 ms · p99 3951 ms
Universal Streamingp50 1481 ms · p99 3261 ms
STT 1p50 1489 ms · p99 2635 ms
Defaultvia Speechmaticsp50 1533 ms · p99 3279 ms
STT RT v5p50 1555 ms · p99 3259 ms
Enhancedp50 1565 ms · p99 3063 ms
Velma 2 STT Streamingp50 1573 ms · p99 3380 ms
Solaria 1p50 1598 ms · p99 6630 ms
Defaultvia Azurep50 1754 ms · p99 3356 ms
Ink 2p50 1773 ms · p99 3558 ms
Voxtral Mini Transcribe Realtime 2602p50 1835 ms · p99 3734 ms
GPT Realtime Whisperp50 1842 ms · p99 3632 ms
Defaultvia Gradiump50 1965 ms · p99 3741 ms
Scribe v2 Realtimep50 2112 ms · p99 3226 ms
Pulsep50 2291 ms · p99 3518 ms
Chirp 3p50 6370 ms · p99 9047 ms
Chirp 2p50 6385 ms · p99 8905 ms

band p25–p75 · tick p50 · whisker to p99 with p90 and p95 stops

Last 30 days

Daily median for the current top 5 · gaps are days without qualifying runs.

Time to First Token — current leadersDaily p50 per model · UTC days

FluxFlux MultilingualParakeet TDT 0.6B v3Universal 3.5 ProWhisper Large v3

Time to First Token by dataset

The overall average split by test condition — where each model holds up and where it degrades.

Time to First Token for every STT model, per dataset, over the last 30 days.
#	Model	Host	All datasets	WildASR clean	PipeCat (production)	WildASR accents	WildASR clipping	WildASR far-field	WildASR noise gaps	WildASR phone codec	WildASR reverb
1	Universal 3.5 Pro	AssemblyAI	1030 ms	899 ms	1181 ms	464 ms	927 ms	959 ms	916 ms	956 ms	923 ms
2	Flux	Deepgram	1089 ms	931 ms	1281 ms	703 ms	838 ms	828 ms	930 ms	1044 ms	826 ms
3	Flux Multilingual	Deepgram	1175 ms	1008 ms	1348 ms	477 ms	1216 ms	1103 ms	1021 ms	1035 ms	1075 ms
4	Parakeet TDT 0.6B v3	Together AI	1210 ms	1111 ms	1365 ms	489 ms	1088 ms	1112 ms	1135 ms	1133 ms	1099 ms
5	Whisper Large v3	Together AI	1241 ms	1150 ms	1399 ms	452 ms	1104 ms	1144 ms	1162 ms	1173 ms	1146 ms
6	Nova 3	Deepgram	1434 ms	1225 ms	1641 ms	989 ms	1289 ms	1268 ms	1242 ms	1236 ms	1263 ms
7	Nova 2	Deepgram	1436 ms	1211 ms	1631 ms	1024 ms	1416 ms	1301 ms	1230 ms	1267 ms	1264 ms
8	STT 1	Inworld AI	1468 ms	1432 ms	1534 ms	1073 ms	1328 ms	1454 ms	1436 ms	1552 ms	1415 ms
9	Default	Speechmatics	1473 ms	1393 ms	1575 ms	715 ms	1540 ms	1492 ms	1436 ms	1448 ms	1462 ms
10	Universal Streaming	AssemblyAI	1510 ms	1376 ms	1616 ms	866 ms	1661 ms	1560 ms	1418 ms	1517 ms	1469 ms
11	STT RT v5	Soniox	1533 ms	1447 ms	1655 ms	817 ms	1569 ms	1514 ms	1469 ms	1455 ms	1457 ms
12	Enhanced	Speechmatics	1536 ms	1467 ms	1635 ms	861 ms	1562 ms	1555 ms	1486 ms	1498 ms	1502 ms
13	Nemotron 3.5 ASR Streaming	Together AI	1540 ms	1378 ms	1648 ms	937 ms	1900 ms	1588 ms	1432 ms	1408 ms	1499 ms
14	Velma 2 STT Streaming	Modulate	1576 ms	1491 ms	1697 ms	868 ms	1591 ms	1542 ms	1504 ms	1478 ms	1523 ms
15	Solaria 1	Gladia	1703 ms	1568 ms	1839 ms	1017 ms	1749 ms	1618 ms	1642 ms	1680 ms	1611 ms
16	Default	Azure	1792 ms	1702 ms	1860 ms	1094 ms	2051 ms	1880 ms	1787 ms	1828 ms	1784 ms
17	Ink 2	Cartesia	1812 ms	1763 ms	1921 ms	1024 ms	1790 ms	1814 ms	1778 ms	1764 ms	1774 ms
18	GPT Realtime Whisper	OpenAI	1823 ms	1767 ms	1923 ms	1043 ms	1845 ms	1855 ms	1781 ms	1790 ms	1815 ms
19	Voxtral Mini Transcribe Realtime 2602	Mistral	1828 ms	1774 ms	1937 ms	1050 ms	1806 ms	1819 ms	1795 ms	1792 ms	1802 ms
20	Default	Gradium	1979 ms	1926 ms	2094 ms	1262 ms	1882 ms	1923 ms	1953 ms	1936 ms	1931 ms
21	Pulse	Smallest	2072 ms	1933 ms	2211 ms	1337 ms	2175 ms	2121 ms	1945 ms	2006 ms	1954 ms
22	Scribe v2 Realtime	ElevenLabs	2157 ms	2114 ms	2199 ms	2073 ms	2140 ms	2125 ms	2119 ms	2114 ms	2117 ms
23	Chirp 3	Google	5999 ms	6215 ms	5934 ms	4451 ms	6381 ms	6301 ms	6293 ms	6392 ms	5977 ms
24	Chirp 2	Google	6000 ms	6211 ms	5933 ms	4455 ms	6397 ms	6317 ms	6324 ms	6377 ms	5970 ms

Ranked on the all-datasets average; a dash means the model was not measured on that dataset in the last 30 days. Cell shading deepens toward each column’s highest value.

About TTFT

Why it matters

Partial transcripts support live captions and allow an application to begin processing speech before the turn is complete. TTFT measures when the first partial transcript arrives.

TTFT should be read with Time to Final Segment because an early partial transcript does not guarantee an early final transcript.

How Coval measures it

The same audio runs against every model daily, with connection setup (TCP, TLS, handshakes) excluded the same way for everyone.

Some providers emit partial transcripts on a fixed schedule rather than as soon as text is available. Those endpoints are excluded from TTFT rankings because the result would primarily reflect the emission interval; their other metrics remain available.

Caveats and interpretation

Providers that emit partials on a fixed schedule can expose their event cadence as much as model computation time.
A fast first token does not guarantee a fast or stable final transcript, so TTFT should be read beside TTFS and WER.

The datasets behind it

Every model runs the same fixed inputs, so a gap in TTFT is the model's doing — not the test's.

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