BENCHMARKSTT TTS58 MODELS RANKEDLAST 30 DAYS

Word Error Rate (WER)

Name: Word Error Rate voice AI benchmark results
Creator: Coval

Word Error Rate (WER) is the share of words a transcript gets wrong: substitutions, insertions and deletions, measured against a reference.

Current STT leader#1 / 28: 3.2%; Universal 3.5 Provia AssemblyAI
Current TTS leader#1 / 30: 3.7%; TTS RT v1via Soniox

How it is calculated

WER = (substitutions + deletions + insertions) ÷ reference-word count × 100.

How to read it

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

Speech-to-Text models on Word Error Rate

Full STT dashboard

Word Error Rate — every measured STT modelPercent · lower is better · 30-day average

Show all 28 models

#13Defaultvia Azure5.3%
#14Defaultvia Speechmatics5.5%
#15Scribe v2 Realtime5.5%
#16STT RT v55.8%
#17Voxtral Mini Transcribe Realtime 26026.2%
#18Nova 36.3%
#19Velma 2 STT Streaming6.4%
#20Flux6.6%
#21Universal Streaming7.0%
#22Solaria 17.1%
#23Nova 27.9%
#24Whisper Large v38.2%
#25Flux Multilingual8.6%
#26Defaultvia Gradium10.0%
#27Parakeet TDT 0.6B v311.0%
#28Nemotron 3.5 ASR Streaming15.7%

median of all models · 5.5%

STT models over the last 30 days, ranked on Word Error Rate.
#	Model	Host	TTFS	TTFT	Samples
1	Universal 3.5 Pro	AssemblyAI	146 ms	1030 ms	29,315
2	resonant-1	Reson8	288 ms	—	17,091
3	Chirp 3	Google	813 ms	5999 ms	29,360
4	Enhanced	Speechmatics	341 ms	1536 ms	29,356
5	GPT-4o Transcribe	OpenAI	742 ms	—	29,359
6	Grok STT	xAI	199 ms	—	29,363
7	STT 1	Inworld AI	83 ms	1468 ms	29,321
8	Ink 2	Cartesia	108 ms	1812 ms	29,329
9	Chirp 2	Google	811 ms	6000 ms	29,278
10	GPT-4o mini Transcribe	OpenAI	623 ms	—	29,359
11	GPT Realtime Whisper	OpenAI	559 ms	1823 ms	29,327
12	Pulse	Smallest	205 ms	2072 ms	29,354
13	Default	Azure	155 ms	1792 ms	6,689
14	Default	Speechmatics	223 ms	1473 ms	29,353
15	Scribe v2 Realtime	ElevenLabs	120 ms	2157 ms	29,354
16	STT RT v5	Soniox	64 ms	1533 ms	29,363
17	Voxtral Mini Transcribe Realtime 2602	Mistral	356 ms	1828 ms	29,206
18	Nova 3	Deepgram	99 ms	1434 ms	29,335
19	Velma 2 STT Streaming	Modulate	191 ms	1576 ms	17,919
20	Flux	Deepgram	—	1089 ms	29,354
21	Universal Streaming	AssemblyAI	—	1510 ms	29,297
22	Solaria 1	Gladia	680 ms	1703 ms	26,391
23	Nova 2	Deepgram	101 ms	1436 ms	29,310
24	Whisper Large v3	Together AI	180 ms	1241 ms	29,232
25	Flux Multilingual	Deepgram	—	1175 ms	29,344
26	Default	Gradium	249 ms	1979 ms	28,273
27	Parakeet TDT 0.6B v3	Together AI	70 ms	1210 ms	29,201
28	Nemotron 3.5 ASR Streaming	Together AI	—	1540 ms	29,193

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

Speech-to-Text: where the errors come from

WER composition — STTPercentage points of average WER · lower is better

Universal 3.5 Pro3.2%
resonant-13.4%
Chirp 34.1%
Enhanced4.3%
GPT-4o Transcribe4.7%
Grok STT4.8%
STT 14.8%
Ink 24.9%
Chirp 25.0%
GPT-4o mini Transcribe5.1%
GPT Realtime Whisper5.2%
Pulse5.3%
Defaultvia Azure5.3%
Defaultvia Speechmatics5.5%
Scribe v2 Realtime5.5%
STT RT v55.8%
Voxtral Mini Transcribe Realtime 26026.2%
Nova 36.3%
Velma 2 STT Streaming6.4%
Flux6.6%
Universal Streaming7.0%
Solaria 17.1%
Nova 27.9%
Whisper Large v38.2%
Flux Multilingual8.6%
Defaultvia Gradium10.0%
Parakeet TDT 0.6B v311.0%
Nemotron 3.5 ASR Streaming15.7%

SubstitutionsDeletionsInsertions

Models without a published error split are not shown here.

Speech-to-Text: last 30 days

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

Word Error Rate — current leadersDaily average per model · UTC days

Chirp 3EnhancedGPT-4o Transcriberesonant-1Universal 3.5 Pro

Speech-to-Text: Word Error Rate by dataset

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

Word Error Rate for every STT model, per dataset, over the last 30 days.
#	Model	Host
1	Universal 3.5 Pro	AssemblyAI
2	resonant-1	Reson8
3	Chirp 3	Google
4	Enhanced	Speechmatics
5	GPT-4o Transcribe	OpenAI
6	Grok STT	xAI
7	STT 1	Inworld AI
8	Ink 2	Cartesia
9	Chirp 2	Google
10	GPT-4o mini Transcribe	OpenAI
11	GPT Realtime Whisper	OpenAI
12	Pulse	Smallest
13	Default	Azure
14	Default	Speechmatics
15	Scribe v2 Realtime	ElevenLabs
16	STT RT v5	Soniox
17	Voxtral Mini Transcribe Realtime 2602	Mistral
18	Nova 3	Deepgram
19	Velma 2 STT Streaming	Modulate
20	Flux	Deepgram
21	Universal Streaming	AssemblyAI
22	Solaria 1	Gladia
23	Nova 2	Deepgram
24	Whisper Large v3	Together AI
25	Flux Multilingual	Deepgram
26	Default	Gradium
27	Parakeet TDT 0.6B v3	Together AI
28	Nemotron 3.5 ASR Streaming	Together AI

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. Dotted values split into substitutions, deletions and insertions on hover or tap.

Text-to-Speech models on Word Error Rate

Full TTS dashboard

Word Error Rate — every measured TTS modelPercent · lower is better · 30-day average

#1TTS RT v13.7%
#2TTS Rt v23.9%
#3Neural4.3%
#4Eleven v3 Conversational4.4%
#5Lightning v3.1 Pro4.4%
#6Default4.6%
#7S2.1 Pro4.6%
#8S2.1 Pro Free4.7%
#9Simba 3.24.7%
#10Speech 2.8 HD4.7%
#11Grok TTS4.7%
#12GPT-4o mini TTS4.8%

Show all 30 models

#13TTS 24.8%
#14Speech 2.8 Turbo4.8%
#15TTS Flash 24.9%
#16S14.9%
#17Dragon HD Latest5.1%
#18Chirp 3 HD5.2%
#19Coda5.2%
#20Falcon 25.2%
#21Simba 3.05.3%
#22Aura 25.3%
#23Sonic 3.65.5%
#24Palabra TTS v15.9%
#25Sonic 3.56.1%
#26Mist v36.3%
#27Flash v2.56.7%
#28Blizzard7.4%
#29vui7.9%
#30Qwen3 TTS Flash Realtime8.8%

median of all models · 4.9%

TTS models over the last 30 days, ranked on Word Error Rate.
#	Model	Host	TTFA	WER	Samples
1	TTS RT v1	Soniox	275 ms	3.7%	14,528
2	TTS Rt v2	Soniox	262 ms		8,849
3	Neural	Azure	236 ms	4.3%	3,350
4	Eleven v3 Conversational	ElevenLabs	412 ms		6,750
5	Lightning v3.1 Pro	Smallest	586 ms	4.4%	14,515
6	Default	Gradium	384 ms	4.6%	14,002
7	S2.1 Pro	Fish Audio	374 ms		10,222
8	S2.1 Pro Free	Fish Audio	806 ms	4.7%	14,464
9	Simba 3.2	Speechify	484 ms	4.7%	14,525
10	Speech 2.8 HD	MiniMax	460 ms		1,995
11	Grok TTS	xAI	420 ms	4.7%	14,524
12	GPT-4o mini TTS	OpenAI	1075 ms	4.8%	14,527
13	TTS 2	Inworld AI	176 ms	4.8%	14,529
14	Speech 2.8 Turbo	MiniMax	411 ms	4.8%	1,997
15	TTS Flash 2	Inworld AI	128 ms		8,719
16	S1	Fish Audio	434 ms	4.9%	10,236
17	Dragon HD Latest	Azure	310 ms	5.1%	3,350
18	Chirp 3 HD	Google	512 ms	5.2%	14,530
19	Coda	Rime	313 ms	5.2%	14,523
20	Falcon 2	Murf	549 ms		10,229
21	Simba 3.0	Speechify	526 ms	5.3%	14,529
22	Aura 2	Deepgram	328 ms	5.3%	14,501
23	Sonic 3.6	Cartesia	466 ms		560
24	Palabra TTS v1	Palabra	116 ms	5.9%	14,396
25	Sonic 3.5	Cartesia	274 ms	6.1%	14,513
26	Mist v3	Rime	256 ms	6.3%	14,524
27	Flash v2.5	ElevenLabs	455 ms	6.7%	9,259
28	Blizzard	Lmnt	235 ms	7.4%	14,117
29	vui	Fluxions	124 ms		8,624
30	Qwen3 TTS Flash Realtime	Alibaba	645 ms	8.8%	14,530

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

Text-to-Speech: where the errors come from

WER composition — TTSPercentage points of average WER · lower is better

SubstitutionsDeletionsInsertions

Models without a published error split are not shown here.

Text-to-Speech: last 30 days

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

Word Error Rate — current leadersDaily average per model · UTC days

Eleven v3 ConversationalLightning v3.1 ProNeuralTTS RT v1TTS Rt v2

About WER

Why it matters

Transcription errors can change names, numbers and intent before the text reaches the rest of a voice application. WER measures these errors at the word level.

Coval reports WER across clean speech, accents, intermittent noise, reverb, microphone distance, clipping, phone codecs and spontaneous speech because recognition accuracy varies by audio condition.

How Coval measures it

For speech-to-text, both the reference and the model transcript pass through Whisper's EnglishTextNormalizer before scoring. This prevents formatting differences, such as a written number versus the same number spelled out, from counting as recognition errors.

For text-to-speech, one fixed ASR model (OpenAI whisper-1) transcribes the generated audio. That transcript is compared with the input text as an intelligibility check. Errors from the fixed transcriber affect every TTS provider.

Caveats and interpretation

Reference quality and text normalization set a floor on absolute WER. Coval applies the same references and normalizer to every model in a category.
An overall WER can hide differences between clean speech, phone codecs and room acoustics. Dataset-level results show performance for each condition.

The datasets behind it

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