Hayz Labs · https://hayzlabs.com/reviews/top/speech-to-text/
reviews · top 20 list · checked 18 Oct
Top AI speech-to-text models
Models that turn recordings into text: meetings, interviews, dictation and subtitles.
Nothing on this list is scored yet: no outside source we can turn into a number covers them. The order below follows the public measure named on each row.
Ordered by an independent word error rate test (fewer mistakes first). One row per model line, newest version only. Open the Table to see the number that placed each one and how it was scored.
- 1Fun-ASR-FlashAlibaba
Lowest error rate on the board; for teams able to use Alibaba Cloud Model Studio
Best for Teams on Alibaba Cloud Model Studio who want the fewest transcript errorsYou give up Price not in our sources yet; the board lists it under a preview name
Not scored yetNo public price - 2MAI-Transcribe-2Microsoft
Among the most accurate transcription models at $0.10 an hour. Still a preview with no SLA, so not for production yet.
Best for Developers transcribing lots of recorded audio who can accept a previewYou give up Preview with no service guarantee; price after 2026 not announced
Not scored yetPay as you go - 3Scribe v2ElevenLabs
Very accurate transcription with speaker labels and a free plan. Heavy API users pay less with MAI-Transcribe-2.
Best for Podcasters and journalists who want speaker labels, with a free way to startYou give up Over twice MAI-Transcribe-2's hourly price; trains on your audio unless you opt out
Not scored yetFree - 4Grok Voice Transcribe 2.0SpaceXAI
Near-top accuracy at the lowest price per hour among the leaders
Best for Large batch or live transcription jobs where cost per hour matters mostYou give up Same batch price as MAI-Transcribe-2 but ranks below it; live costs twice as much
Not scored yetPay as you go - 5StepAudio 3 ASRStepFun
Accurate transcription from StepFun; price not confirmed on StepFun's site
Best for Accurate transcription for teams willing to work with StepFun directlyYou give up Ranks below MAI-Transcribe-2 and Grok, which list $0.10 an hour; its price is unclear
Not scored yetNo public price - 6Gemini 3.5 TranscribeGoogle
Google's transcription model with a free tier, easy if you already use Gemini
Best for Gemini users who want transcription with a free tier and a live streaming versionYou give up Ranks below Scribe v2, which costs less per hour; batch price only from a third party
Not scored yetFree - 7Voxtral SmallMistral AI
The most accurate open speech-to-text model on this board; you can run it yourself.
Best for Running transcription on your own hardware under a permissive Apache licenceYou give up Seven models on the board make fewer errors; you need your own server to run it
Not scored yetFree
| # | Name | Score | What placed it | Price | Best for | You give up |
|---|---|---|---|---|---|---|
| 1 | Fun-ASR-FlashAlibaba · Closed | Not scored | About 2 words in 100 wrong, #1 in a transcription test (Artificial Analysis)NotesTheir figure: AA WER 1.7% (listed as Fun-Realtime-ASR-preview). AA speech-to-text WER #1 of distinct models (lower is better; ties take lowest rank); no second eval to adjust with Report a mistake | Not confirmed | Teams on Alibaba Cloud Model Studio who want the fewest transcript errors | Price not in our sources yet; the board lists it under a preview name |
| 2 | MAI-Transcribe-2Microsoft · Closed · our review | Not scored | About 2 words in 100 wrong, #2 in a transcription test (Artificial Analysis)NotesTheir figure: AA WER 2.0%. AA speech-to-text WER #2 of distinct models (lower is better; ties take lowest rank); no second eval to adjust with; matches review Report a mistake | API $0.10 per hour of audio (introductory) | Developers transcribing lots of recorded audio who can accept a preview | Preview with no service guarantee; price after 2026 not announced |
| 3 | Scribe v2ElevenLabs · Closed · our review | Not scored | About 2 words in 100 wrong, #3 in a transcription test (Artificial Analysis)NotesTheir figure: AA WER 2.2%. AA speech-to-text WER #3 of distinct models (lower is better; ties take lowest rank); no second eval to adjust with; board moved it to #3. Report a mistake | API $0.22 per hour of audio; Starter $6/mo | Podcasters and journalists who want speaker labels, with a free way to start | Over twice MAI-Transcribe-2's hourly price; trains on your audio unless you opt out |
| 4 | Grok Voice Transcribe 2.0SpaceXAI · Closed | Not scored | About 2 words in 100 wrong, #4 in a transcription test (Artificial Analysis)NotesTheir figure: AA WER 2.3%. AA speech-to-text WER #4 of distinct models (lower is better; ties take lowest rank); no second eval to adjust with Report a mistake | API $0.10 per hour batch, $0.20 per hour streaming | Large batch or live transcription jobs where cost per hour matters most | Same batch price as MAI-Transcribe-2 but ranks below it; live costs twice as much |
| 5 | StepAudio 3 ASRStepFun · Closed | Not scored | About 3 words in 100 wrong, #6 in a transcription test (Artificial Analysis)NotesTheir figure: AA WER 2.5%. AA speech-to-text WER #6 of distinct models (lower is better; ties take lowest rank); no second eval to adjust with; MAI-Transcribe-1.5 holds #5 but is an older line Report a mistake | Not confirmed | Accurate transcription for teams willing to work with StepFun directly | Ranks below MAI-Transcribe-2 and Grok, which list $0.10 an hour; its price is unclear |
| 6 | Gemini 3.5 TranscribeGoogle · Closed | Not scored | About 3 words in 100 wrong, #7 in a transcription test (Artificial Analysis)NotesTheir figure: AA WER 2.6%. AA speech-to-text WER #7 of distinct models (lower is better; ties take lowest rank); no second eval to adjust with Report a mistake | Free tier; API about $0.005 per minute ($2 in / $12 out per 1M tokens) | Gemini users who want transcription with a free tier and a live streaming version | Ranks below Scribe v2, which costs less per hour; batch price only from a third party |
| 7 | Voxtral SmallMistral AI · Open weights | Not scored | About 3 words in 100 wrong, #10 in a transcription test (Artificial Analysis)NotesTheir figure: AA WER 2.8%. AA speech-to-text WER #10 of distinct models (tied at 2.8% with Gemini 3.1 Pro Preview (High) and Zoom Scribe V2 Pro; 7 distinct models score better; ties take the lowest rank); no second eval to adjust with Report a mistake | Free (open weights) | Running transcription on your own hardware under a permissive Apache licence | Seven models on the board make fewer errors; you need your own server to run it |
Nothing matches all of that. , or try everything.
How we ranked this
Order. Ordered by an independent word error rate test (fewer mistakes first). One row per model line, newest version only.
How we score this list. No score yet: we have not set a rule for turning a word error rate into a score. The order is the test's. The full rules are on How we score. Open "How it was scored" on any row to see the working.
What we did not do. We did not run any of these ourselves. Every number comes from the public source linked on its row, checked on 18 Oct 2026.
Spot a mistake? A wrong rank, price or link: tell us and we will check it against the source. Corrections are dated in News.
Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule.