BENCHMARKSTT27 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 / 27: 924ms; Whisper Large v3via Baseten

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

#1Whisper Large v3via Baseten924 ms
#2Qwen3 ASR 1.7b946 ms
#3Universal 3.5 Pro1026 ms
#4Universal 3.6 Pro1072 ms
#5Flux1091 ms
#6Flux Multilingual1165 ms
#7STT 11207 ms
#8Parakeet TDT 0.6B v31228 ms
#9Whisper Large v3via Together AI1260 ms
#10Nova 21417 ms
#11Nova 31432 ms
#12Universal Streaming1509 ms

Show all 27 models

#13STT RT v51518 ms
#14Linden 11537 ms
#15Nemotron 3.5 ASR Streaming1544 ms
#16Gemini 3.5 Transcribe Live1604 ms
#17Qwen3 ASR Fast1752 ms
#18Solaria 11772 ms
#19GPT Realtime Whisper1824 ms
#20Voxtral Mini Transcribe Realtime 26021852 ms
#21Ink 21968 ms
#22Pulse1976 ms
#23Default1986 ms
#24Scribe v2 Realtime2157 ms
#25Daytona STT5695 ms
#26Chirp 35956 ms
#27Chirp 26099 ms

median of all models · 1537 ms

STT models over the last 30 days, ranked on Time to First Token.
#	Model	Host	TTFS	TTFT	Samples
1	Whisper Large v3	Baseten	131 ms	924 ms	13,663
2	Qwen3 ASR 1.7b	Baseten	39 ms	946 ms	13,669
3	Universal 3.5 Pro	AssemblyAI	169 ms	1026 ms	26,654
4	Universal 3.6 Pro	AssemblyAI	135 ms	1072 ms	24,470
5	Flux	Deepgram	99 ms	1091 ms	26,005
6	Flux Multilingual	Deepgram	98 ms	1165 ms	26,007
7	STT 1	Inworld AI	67 ms	1207 ms	24,749
8	Parakeet TDT 0.6B v3	Together AI	97 ms	1228 ms	26,439
9	Whisper Large v3	Together AI	199 ms	1260 ms	26,390
10	Nova 2	Deepgram	87 ms	1417 ms	26,002
11	Nova 3	Deepgram	85 ms	1432 ms	26,005
12	Universal Streaming	AssemblyAI	—	1509 ms	26,626
13	STT RT v5	Soniox	45 ms	1518 ms	23,533
14	Linden 1	Speechmatics	272 ms	1537 ms	26,595
15	Nemotron 3.5 ASR Streaming	Together AI	—	1544 ms	26,549
16	Gemini 3.5 Transcribe Live	Gemini	317 ms	1604 ms	26,578
17	Qwen3 ASR Fast	Nari Labs	47 ms	1752 ms	26,652
18	Solaria 1	Gladia	623 ms	1772 ms	26,240
19	GPT Realtime Whisper	OpenAI	599 ms	1824 ms	26,622
20	Voxtral Mini Transcribe Realtime 2602	Mistral	412 ms	1852 ms	26,109
21	Ink 2	Cartesia	149 ms	1968 ms	26,653
22	Pulse	Smallest	205 ms	1976 ms	26,654
23	Default	Gradium	260 ms	1986 ms	26,632
24	Scribe v2 Realtime	ElevenLabs	122 ms	2157 ms	26,624
25	Daytona STT	Guava	399 ms	5695 ms	5,225
26	Chirp 3	Google	768 ms	5956 ms	26,638
27	Chirp 2	Google	909 ms	6099 ms	26,650
—	GPT-4o mini Transcribe	OpenAI	709 ms	—	26,621
—	GPT-4o Transcribe	OpenAI	784 ms	—	26,625
—	Grok STT	SpaceXAI	223 ms	—	26,577
—	resonant-1	Reson8	252 ms	—	26,653

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 993 ms · Max 26984 ms
Nova 3p50 997 ms · Max 5435 ms
Whisper Large v3via Basetenp50 1003 ms · Max 4862 ms
STT 1p50 1046 ms · Max 13056 ms
Qwen3 ASR 1.7bp50 1054 ms · Max 4883 ms
Universal 3.5 Prop50 1117 ms · Max 4818 ms
Universal 3.6 Prop50 1118 ms · Max 5521 ms
Fluxp50 1171 ms · Max 4623 ms
Flux Multilingualp50 1176 ms · Max 5618 ms
Whisper Large v3via Together AIp50 1301 ms · Max 17738 ms
Parakeet TDT 0.6B v3p50 1315 ms · Max 26520 ms
Nemotron 3.5 ASR Streamingp50 1468 ms · Max 9564 ms
Universal Streamingp50 1480 ms · Max 10946 ms
Gemini 3.5 Transcribe Livep50 1526 ms · Max 14199 ms
STT RT v5p50 1535 ms · Max 5483 ms
Linden 1p50 1539 ms · Max 15599 ms
Solaria 1p50 1599 ms · Max 26400 ms
Qwen3 ASR Fastp50 1749 ms · Max 4543 ms
Voxtral Mini Transcribe Realtime 2602p50 1843 ms · Max 36932 ms
GPT Realtime Whisperp50 1880 ms · Max 12852 ms
Defaultp50 1964 ms · Max 7975 ms
Ink 2p50 1993 ms · Max 5656 ms
Pulsep50 2087 ms · Max 5555 ms
Scribe v2 Realtimep50 2112 ms · Max 5785 ms
Daytona STTp50 5354 ms · Max 14582 ms
Chirp 3p50 6342 ms · Max 25530 ms
Chirp 2p50 6445 ms · Max 24490 ms

band p25–p75 · tick p50 · whisker to max 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

FluxQwen3 ASR 1.7bUniversal 3.5 ProUniversal 3.6 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
1	Whisper Large v3	Baseten	924 ms	819 ms	1048 ms	417 ms	835 ms	861 ms	835 ms	848 ms
2	Qwen3 ASR 1.7b	Baseten	946 ms	830 ms	1079 ms	280 ms	904 ms	887 ms	863 ms	925 ms
3	Universal 3.5 Pro	AssemblyAI	1026 ms	898 ms	1172 ms	440 ms	925 ms	950 ms	904 ms	955 ms
4	Universal 3.6 Pro	AssemblyAI	1072 ms	951 ms	1206 ms	443 ms	1027 ms	1085 ms	962 ms	985 ms
5	Flux	Deepgram	1091 ms	923 ms	1275 ms	689 ms	825 ms	809 ms	930 ms	1040 ms
6	Flux Multilingual	Deepgram	1165 ms	1004 ms	1334 ms	459 ms	1205 ms	1088 ms	1011 ms	1025 ms
7	STT 1	Inworld AI	1207 ms	1162 ms	1255 ms	1048 ms	1122 ms	1178 ms	1165 ms	1219 ms
8	Parakeet TDT 0.6B v3	Together AI	1228 ms	1140 ms	1371 ms	508 ms	1107 ms	1144 ms	1141 ms	1161 ms
9	Whisper Large v3	Together AI	1260 ms	1185 ms	1397 ms	488 ms	1160 ms	1188 ms	1192 ms	1207 ms
10	Nova 2	Deepgram	1417 ms	1201 ms	1601 ms	1004 ms	1362 ms	1315 ms	1204 ms	1233 ms
11	Nova 3	Deepgram	1432 ms	1215 ms	1640 ms	970 ms	1272 ms	1270 ms	1223 ms	1215 ms
12	Universal Streaming	AssemblyAI	1509 ms	1388 ms	1611 ms	848 ms	1655 ms	1562 ms	1423 ms	1518 ms
13	STT RT v5	Soniox	1518 ms	1442 ms	1631 ms	804 ms	1554 ms	1506 ms	1455 ms	1441 ms
14	Linden 1	Speechmatics	1537 ms	1478 ms	1626 ms	832 ms	1575 ms	1604 ms	1497 ms	1520 ms
15	Nemotron 3.5 ASR Streaming	Together AI	1544 ms	1391 ms	1644 ms	950 ms	1965 ms	1596 ms	1427 ms	1405 ms
16	Gemini 3.5 Transcribe Live	Gemini	1604 ms	1571 ms	1656 ms	834 ms	1756 ms	1714 ms	1598 ms	1732 ms
17	Qwen3 ASR Fast	Nari Labs	1752 ms	1752 ms	1751 ms	1757 ms	1751 ms	1751 ms	1750 ms	1751 ms
18	Solaria 1	Gladia	1772 ms	1641 ms	1909 ms	1164 ms	1760 ms	1718 ms	1646 ms	1728 ms
19	GPT Realtime Whisper	OpenAI	1824 ms	1776 ms	1921 ms	1052 ms	1855 ms	1852 ms	1787 ms	1794 ms
20	Voxtral Mini Transcribe Realtime 2602	Mistral	1852 ms	1806 ms	1963 ms	1070 ms	1820 ms	1833 ms	1819 ms	1795 ms
21	Ink 2	Cartesia	1968 ms	1921 ms	2077 ms	1193 ms	1924 ms	1967 ms	1927 ms	1923 ms
22	Pulse	Smallest	1976 ms	1900 ms	2075 ms	1198 ms	2095 ms	2022 ms	1909 ms	1970 ms
23	Default	Gradium	1986 ms	1939 ms	2096 ms	1245 ms	1902 ms	1945 ms	1968 ms	1959 ms
24	Scribe v2 Realtime	ElevenLabs	2157 ms	2116 ms	2195 ms	2068 ms	2133 ms	2129 ms	2127 ms	2121 ms
25	Daytona STT	Guava	5695 ms	5663 ms	5928 ms	3529 ms	5956 ms	6508 ms	4570 ms	5750 ms
26	Chirp 3	Google	5956 ms	6180 ms	5878 ms	4426 ms	6341 ms	6297 ms	6268 ms	6329 ms
27	Chirp 2	Google	6099 ms	6335 ms	6008 ms	4562 ms	6499 ms	6492 ms	6399 ms	6503 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