Which speech-to-text engine is actually the best?

Twenty engines, one accuracy source, and the five things that decide whether a transcript is usable rather than merely cheap: what it misses, what it costs, how long it takes, whether it separates the voices, and whether it knows your vocabulary. Nobody here is paying us to be in this table.

Accuracy is Artificial Analysis’s, an index that runs the same files through 55 models and sells none of them. Vendors’ own numbers are left out: they contradict each other. Prices are dollars per hour of audio with the glossary switched on where it costs extra. Checked 7 September 2026.

What a percentage means in a real meeting

A ninety-minute meeting is about 13,500 words. That is the difference between an accuracy number and a decision: the best engine gets 270 of them wrong, and the cheapest one that still separates voices gets 630.

Most accurate

Microsoft MAI-Transcribe-2

270

wrong words in one meeting

Cheapest

StepFun StepAudio 2.5 ASR

630

wrong words in one meeting

Engine	Errors	An hour	Speed	Speakers	Word times	Glossary
Fun-Realtime-ASRAlibabathe best score on the market, and it is only measured on the streaming board: absent from the file one. The streaming model does not separate the voices and takes 32 terms	Errors1.7%230	An hourno tariff	SpeedNot yet measured	SpeakersNo	Word timesYes	Glossary32
MAI-Transcribe-2Microsoftthe price is promotional and the vendor has published its end date, 31 December 2026	Errors2.0%270	An hour$0.10	Speed411×	SpeakersYes	Word timesYes	Glossary1,000
Scribe v2ElevenLabsIn Scribe	Errors2.2%300	An hour$0.26	Speed52×	SpeakersYes	Word timesYes	Glossary1,000
Pulse ProSmallest AIEnglish only, and no documented glossary	Errors2.4%320	An hour$0.24	Speed262×	SpeakersYes	Word timesYes	GlossaryNone
MAI-Transcribe-1.5Microsoftreplaced by MAI-Transcribe-2, more accurate and three times cheaper	Errors2.4%320	An hour$0.36	Speed195×	SpeakersYes	Word timesYes	Glossary1,000
Gemini 3.5 TranscribeGooglethirty minutes per request as soon as the voices are separated, so a long meeting gets cut up and the voices stop matching from one cut to the next	Errors2.6%350	An hour$0.30	SpeedNot yet measured	SpeakersYes	Word timesYes	Glossary1,000
Voxtral SmallMistralan instruction model you talk to, not a transcription engine: it returns a block of text with no voices and no timings, and it costs twice the Voxtral that returns both	Errors2.8%380	An hour$0.36	SpeedNot yet measured	SpeakersNo	Word timesNo	GlossaryNone
Universal-3.5 ProAssemblyAIIn Scribe	Errors3.0%410	An hour$0.28	SpeedNot yet measured	SpeakersYes	Word timesYes	Glossary1,000
Solaria-3Gladiatwice the price of a more accurate engine	Errors3.2%430	An hour$0.61	SpeedNot yet measured	SpeakersYes	Word timesYes	Glossary1,000
GPT TranscribeOpenAIIn Scribeno voices and no word timings: a block of text you cannot click into	Errors3.3%450	An hour$0.27	SpeedNot yet measured	SpeakersNo	Word timesNo	GlossaryNone
Resonant-1Reson8	Errors3.4%460	An hour$0.22	SpeedNot yet measured	SpeakersYes	Word timesYes	GlossaryNone
Inkling (256K)Thinking Machinesno published tariff	Errors3.5%470	An hourno tariff	SpeedNot yet measured	SpeakersNo	Word timesNo	GlossaryNone
Voxtral Mini Transcribe 2MistralIn Scribethirteen languages, and a glossary ten times smaller	Errors3.6%490	An hour$0.18	SpeedNot yet measured	SpeakersYes	Word timesYes	Glossary100
Voxtral MiniMistral via DeepInfrathe cheapest engine that still separates the voices	Errors3.8%510	An hour$0.060	SpeedNot yet measured	SpeakersYes	Word timesYes	Glossary100
GPT-4o TranscribeOpenAIreplaced by the same vendor's newer model, at a higher price	Errors4.0%540	An hour$0.36	SpeedNot yet measured	SpeakersNo	Word timesNo	GlossaryNone
Parakeet TDT 0.6B v3NVIDIA via Togetherno voice separation	Errors4.5%610	An hour$0.090	SpeedNot yet measured	SpeakersNo	Word timesYes	GlossaryNone
Whisper Large v3 TurboOpenAI via Groqno voice separation, and the glossary is a short brief	Errors4.6%620	An hour$0.040	SpeedNot yet measured	SpeakersNo	Word timesYes	GlossaryNone
StepAudio 2.5 ASRStepFunthe cheapest on the market, and it returns nothing but a block of text	Errors4.7%630	An hour$0.022	SpeedNot yet measured	SpeakersNo	Word timesNo	GlossaryNone
MeliaSpeechmatics	Errors4.9%660	An hour$0.24	SpeedNot yet measured	SpeakersYes	Word timesYes	Glossary1,000
Nova-3Deepgramthe fastest on the market, and the least accurate of the big ones. The voices are a paid option	Errors5.2%700	An hour$0.26	Speed606×	SpeakersYes	Word timesYes	Glossary1,000

Two engines launched too recently for the index to have run them. Their vendors claim good numbers; so does everybody. They are listed without a rank rather than ranked on their own word.

What we chose, and why

Scribe runs on the top of this table by default, and lets you drop down it whenever a recording does not deserve the best engine. That is the whole argument for holding your own key: the cheapest row costs a fifteenth of the dearest and gets twice as many words wrong, and only you know which of those matters for the thing you just recorded.

You can switch to it inside Scribe, per recording, on your own key.

See what Scribe does with the transcript