BENCHMARKSTT29 MODELS RANKEDLAST 30 DAYS

Time to Final Segment (TTFS)

Name: Time to Final Segment voice AI benchmark results
Creator: Coval

Time to Final Segment (TTFS) is how long a speech-to-text model takes to deliver its final transcript after the speaker stops talking.

Current STT leader#1 / 29: 39ms; Qwen3 ASR 1.7bvia Baseten

How it is calculated

TTFS = final-transcript timestamp − shared end-of-speech timestamp.

How to read it

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

Speech-to-Text models on Time to Final Segment

Full STT dashboard

Time to Final Segment — every measured STT modelMilliseconds · lower is better · 30-day average

#1Qwen3 ASR 1.7b39 ms
#2STT RT v545 ms
#3Qwen3 ASR Fast47 ms
#4STT 166 ms
#5Nova 385 ms
#6Nova 287 ms
#7Parakeet TDT 0.6B v397 ms
#8Flux Multilingual98 ms
#9Flux99 ms
#10Scribe v2 Realtime122 ms
#11Whisper Large v3via Baseten131 ms
#12Universal 3.6 Pro135 ms

Show all 29 models

#13Ink 2148 ms
#14Universal 3.5 Pro168 ms
#15Whisper Large v3via Together AI200 ms
#16Pulse205 ms
#17Grok STT223 ms
#18resonant-1252 ms
#19Default258 ms
#20Linden 1271 ms
#21Gemini 3.5 Transcribe Live317 ms
#22Daytona STT399 ms
#23Voxtral Mini Transcribe Realtime 2602411 ms
#24GPT Realtime Whisper599 ms
#25Solaria 1638 ms
#26GPT-4o mini Transcribe708 ms
#27Chirp 3768 ms
#28GPT-4o Transcribe783 ms
#29Chirp 2908 ms

median of all models · 200 ms

STT models over the last 30 days, ranked on Time to Final Segment.
#	Model	Host	TTFS	TTFT	Samples
1	Qwen3 ASR 1.7b	Baseten	39 ms	946 ms	13,669
2	STT RT v5	Soniox	45 ms	1518 ms	23,533
3	Qwen3 ASR Fast	Nari Labs	47 ms	1752 ms	26,652
4	STT 1	Inworld AI	66 ms	1212 ms	24,749
5	Nova 3	Deepgram	85 ms	1431 ms	26,002
6	Nova 2	Deepgram	87 ms	1417 ms	25,960
7	Parakeet TDT 0.6B v3	Together AI	97 ms	1228 ms	26,460
8	Flux Multilingual	Deepgram	98 ms	1165 ms	26,007
9	Flux	Deepgram	99 ms	1091 ms	26,005
10	Scribe v2 Realtime	ElevenLabs	122 ms	2158 ms	26,640
11	Whisper Large v3	Baseten	131 ms	924 ms	13,663
12	Universal 3.6 Pro	AssemblyAI	135 ms	1071 ms	24,147
13	Ink 2	Cartesia	148 ms	1965 ms	26,652
14	Universal 3.5 Pro	AssemblyAI	168 ms	1025 ms	26,654
15	Whisper Large v3	Together AI	200 ms	1261 ms	26,445
16	Pulse	Smallest	205 ms	1976 ms	26,654
17	Grok STT	SpaceXAI	223 ms	—	26,576
18	resonant-1	Reson8	252 ms	—	26,653
19	Default	Gradium	258 ms	1986 ms	26,632
20	Linden 1	Speechmatics	271 ms	1537 ms	26,595
21	Gemini 3.5 Transcribe Live	Gemini	317 ms	1609 ms	26,567
22	Daytona STT	Guava	399 ms	5695 ms	5,225
23	Voxtral Mini Transcribe Realtime 2602	Mistral	411 ms	1851 ms	26,199
24	GPT Realtime Whisper	OpenAI	599 ms	1824 ms	26,622
25	Solaria 1	Gladia	638 ms	1782 ms	26,198
26	GPT-4o mini Transcribe	OpenAI	708 ms	—	26,621
27	Chirp 3	Google	768 ms	5956 ms	26,638
28	GPT-4o Transcribe	OpenAI	783 ms	—	26,625
29	Chirp 2	Google	908 ms	6098 ms	26,650
—	Nemotron 3.5 ASR Streaming	Together AI	—	1544 ms	26,542
—	Universal Streaming	AssemblyAI	—	1508 ms	26,624

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 TTFS.

Time to Final Segment distribution per modelMilliseconds · shared axis · last 30 days

Qwen3 ASR 1.7bp50 23 ms · Max 369 ms
STT RT v5p50 39 ms · Max 4373 ms
Qwen3 ASR Fastp50 44 ms · Max 8082 ms
STT 1p50 64 ms · Max 4699 ms
Parakeet TDT 0.6B v3p50 73 ms · Max 12950 ms
Nova 3p50 90 ms · Max 19227 ms
Nova 2p50 91 ms · Max 17906 ms
Fluxp50 98 ms · Max 17313 ms
Flux Multilingualp50 99 ms · Max 10805 ms
Whisper Large v3via Together AIp50 111 ms · Max 12594 ms
Whisper Large v3via Basetenp50 114 ms · Max 10654 ms
Scribe v2 Realtimep50 115 ms · Max 5714 ms
Universal 3.6 Prop50 124 ms · Max 13106 ms
Universal 3.5 Prop50 139 ms · Max 20808 ms
Ink 2p50 146 ms · Max 548 ms
Grok STTp50 192 ms · Max 13563 ms
Pulsep50 192 ms · Max 1172 ms
Linden 1p50 236 ms · Max 21148 ms
Defaultp50 244 ms · Max 29362 ms
resonant-1p50 248 ms · Max 2783 ms
Gemini 3.5 Transcribe Livep50 295 ms · Max 4965 ms
Voxtral Mini Transcribe Realtime 2602p50 335 ms · Max 34706 ms
Solaria 1p50 384 ms · Max 18262 ms
Daytona STTp50 391 ms · Max 1474 ms
GPT Realtime Whisperp50 554 ms · Max 25534 ms
GPT-4o mini Transcribep50 648 ms · Max 28470 ms
GPT-4o Transcribep50 747 ms · Max 30943 ms
Chirp 3p50 758 ms · Max 25641 ms
Chirp 2p50 802 ms · Max 19387 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 Final Segment — current leadersDaily p50 per model · UTC days

Nova 3Qwen3 ASR 1.7bQwen3 ASR FastSTT 1STT RT v5

Time to Final Segment by dataset

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

Time to Final Segment 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	Qwen3 ASR 1.7b	Baseten	39 ms	33 ms	32 ms	24 ms	105 ms	84 ms	39 ms	27 ms
2	STT RT v5	Soniox	45 ms	44 ms	45 ms	46 ms	43 ms	43 ms	44 ms	44 ms
3	Qwen3 ASR Fast	Nari Labs	47 ms	47 ms	47 ms	44 ms	46 ms	46 ms	45 ms	47 ms
4	STT 1	Inworld AI	66 ms	70 ms	65 ms	58 ms	70 ms	71 ms	70 ms	66 ms
5	Nova 3	Deepgram	85 ms	89 ms	83 ms	83 ms	91 ms	88 ms	86 ms	91 ms
6	Nova 2	Deepgram	87 ms	88 ms	86 ms	84 ms	92 ms	89 ms	87 ms	94 ms
7	Parakeet TDT 0.6B v3	Together AI	97 ms	96 ms	99 ms	99 ms	92 ms	95 ms	92 ms	91 ms
8	Flux Multilingual	Deepgram	98 ms	100 ms	94 ms	103 ms	111 ms	103 ms	103 ms	105 ms
9	Flux	Deepgram	99 ms	101 ms	96 ms	102 ms	112 ms	100 ms	101 ms	108 ms
10	Scribe v2 Realtime	ElevenLabs	122 ms	121 ms	121 ms	119 ms	130 ms	124 ms	124 ms	123 ms
11	Whisper Large v3	Baseten	131 ms	128 ms	137 ms	90 ms	137 ms	137 ms	116 ms	132 ms
12	Universal 3.6 Pro	AssemblyAI	135 ms	134 ms	137 ms	123 ms	139 ms	138 ms	133 ms	136 ms
13	Ink 2	Cartesia	148 ms	148 ms	147 ms	151 ms	150 ms	151 ms	149 ms	153 ms
14	Universal 3.5 Pro	AssemblyAI	168 ms	157 ms	163 ms	209 ms	178 ms	188 ms	186 ms	178 ms
15	Whisper Large v3	Together AI	200 ms	197 ms	202 ms	216 ms	209 ms	194 ms	173 ms	197 ms
16	Pulse	Smallest	205 ms	206 ms	202 ms	210 ms	212 ms	214 ms	205 ms	218 ms
17	Grok STT	SpaceXAI	223 ms	226 ms	226 ms	208 ms	236 ms	226 ms	217 ms	195 ms
18	resonant-1	Reson8	252 ms	252 ms	254 ms	241 ms	250 ms	252 ms	251 ms	251 ms
19	Default	Gradium	258 ms	252 ms	242 ms	355 ms	262 ms	279 ms	278 ms	299 ms
20	Linden 1	Speechmatics	271 ms	259 ms	264 ms	312 ms	282 ms	304 ms	283 ms	293 ms
21	Gemini 3.5 Transcribe Live	Gemini	317 ms	315 ms	321 ms	292 ms	308 ms	316 ms	319 ms	315 ms
22	Daytona STT	Guava	399 ms	386 ms	404 ms	438 ms	384 ms	404 ms	382 ms	387 ms
23	Voxtral Mini Transcribe Realtime 2602	Mistral	411 ms	413 ms	409 ms	399 ms	429 ms	406 ms	401 ms	427 ms
24	GPT Realtime Whisper	OpenAI	599 ms	587 ms	592 ms	636 ms	638 ms	599 ms	602 ms	631 ms
25	Solaria 1	Gladia	638 ms	617 ms	643 ms	709 ms	669 ms	639 ms	587 ms	609 ms
26	GPT-4o mini Transcribe	OpenAI	708 ms	707 ms	705 ms	649 ms	729 ms	745 ms	713 ms	731 ms
27	Chirp 3	Google	768 ms	764 ms	779 ms	662 ms	744 ms	774 ms	782 ms	787 ms
28	GPT-4o Transcribe	OpenAI	783 ms	784 ms	790 ms	674 ms	796 ms	790 ms	795 ms	792 ms
29	Chirp 2	Google	908 ms	897 ms	924 ms	793 ms	888 ms	917 ms	885 ms	939 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 TTFS

Why it matters

Applications often wait for a final transcript before acting on a completed turn. A model may stream partial text quickly but take longer to mark the segment final.

Coval uses TTFS as the primary speech-to-text ranking metric because it measures when the final transcript becomes available after speech ends.

How Coval measures it

One shared voice-activity model identifies the end of speech, and every provider's timer starts from that point. Provider-specific endpointing decisions are therefore excluded from the comparison.

Every model receives the same audio, and connection setup is excluded consistently across providers.

Caveats and interpretation

A production agent may add its own endpointing delay before the benchmark's shared end-of-speech instant.
Providers expose different finalization semantics; Coval uses the event treated as final by the published integration for each endpoint.

The datasets behind it

Every model runs the same fixed inputs, so a gap in TTFS 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