English speech → usable text

English transcription APIs: accuracy vs price

The cheapest API is not always the one that makes the fewest mistakes. Below, endpoints with a benchmark are ranked by the word error rate the open Open ASR Leaderboard measured for the exact same model; endpoints without one follow, cheapest first. Your monthly cost sits next to each.

Most accurate5.8% WERScribe v2 · ElevenLabs · $0.22/h
Cheapest with a benchmark$0.031 / hourWhisper Large v3 Turbo · Cloudflare Workers AI · 7.8% WER
Your workloadCalculateMonthly cost for your hours

Accuracy: Open ASR Leaderboard (Apache-2.0), results updated May 22, 2026, checked Oct 17, 2026. Prices from each provider's official pricing page. How we compare.

Ranked for your volume

Transcription monthly cost by provider

APIs, not upload apps

Remove a chip to widen your shortlist.

Audio hours / monthFiles or chunks / month

Costs update as you type.
Equal-sized requests; one audio channel.

Lowest estimate matching your requirements

Whisper at Cloudflare Workers AI

17 priced options · English confirmed · base transcription

$2.70estimated / month

17 options with a full price · 6 more need a closer look (below)

Transcription API cost estimates for your requirements. Sort using the column names.
				Workflow / evidence
Scribe v2ElevenLabs	$22.00	5.8%English average · lower is better	$0.22base rate	Recorded files Sources & billing notes Listed API transcription usage rate; optional keyterm prompting, transcript editing and plan fees are excluded. Price snapshot: 2026-10-31. Tracked dataset. Official pricing ↗ · Feature evidence ↗ Accuracy: Open ASR Leaderboard result for `elevenlabs/scribe_v2`; English meetings 11.9%, long-form 6.2%.
Chirp 3 (Speech-to-Text V2)Google Cloud	$96.00	6.6%English average · lower is better	$0.96base rate	Recorded files Sources & billing notes Standard recognition rate, which Google applies to Chirp 3; first volume tier (up to 500,000 minutes a month), cheaper above. Rounded up per second. Price snapshot: 2026-10-17. Sourced preview; confirm the provider quote. Official pricing ↗ · Feature evidence ↗ Accuracy: Open ASR Leaderboard result for `google/chirp_3`; English meetings 11.8%, long-form 9.6%.
EnhancedSpeechmatics	$38.00	6.9%English average · lower is better	$0.38base rate	Recorded files Sources & billing notes Highest-accuracy pre-recorded rate; speaker diarization included. Price snapshot: 2026-10-17. Sourced preview; confirm the provider quote. Official pricing ↗ · Feature evidence ↗ Accuracy: Open ASR Leaderboard result for `speechmatics/enhanced`; English meetings 14.5%, long-form 7.0%.
Whisper Large v3Groq	$11.10	7.4%English average · lower is better	$0.111base rate	Recorded files Sources & billing notes 10-second minimum billed duration per request. Price snapshot: 2026-10-31. Tracked dataset. Official pricing ↗ · Feature evidence ↗ Accuracy: Open ASR Leaderboard result for `openai/whisper-large-v3`; English meetings 15.9%, long-form 8.7%. Model score; the host's decoding settings can differ.
Whisper Large v3 TurboCloudflare Workers AI	$3.06	7.8%English average · lower is better	$0.031base rate	Recorded files Sources & billing notes Workers AI duration estimate; dataset rounds the provider's per-minute rate. Free allocation is not deducted. Price snapshot: 2026-10-31. Tracked dataset. Official pricing ↗ · Feature evidence ↗ Accuracy: Open ASR Leaderboard result for `openai/whisper-large-v3-turbo`; English meetings 16.1%, long-form 8.7%. Model score; the host's decoding settings can differ.
Whisper Large v3 TurboGroq	$4.00	7.8%English average · lower is better	$0.04base rate	Recorded files Sources & billing notes 10-second minimum billed duration per request. Price snapshot: 2026-10-31. Tracked dataset. Official pricing ↗ · Feature evidence ↗ Accuracy: Open ASR Leaderboard result for `openai/whisper-large-v3-turbo`; English meetings 16.1%, long-form 8.7%. Model score; the host's decoding settings can differ.
WhisperCloudflare Workers AILowest estimate	$2.70	Not benchmarked	$0.027base rate	Recorded files Sources & billing notes Workers AI duration estimate; free allocation and account charges are not deducted. Price snapshot: 2026-10-31. Tracked dataset. Official pricing ↗ · Feature evidence ↗
Melia 1Speechmatics	$12.00	Not benchmarked	$0.12base rate	Recorded files Sources & billing notes Pre-recorded pay-as-you-go rate. Multilingual model; speaker diarization included. Price snapshot: 2026-10-17. Sourced preview; confirm the provider quote. Official pricing ↗ · Feature evidence ↗
Universal-2AssemblyAI	$15.00	Not benchmarked	$0.15base rate	Recorded files Sources & billing notes Pay-as-you-go pre-recorded rate per hour of audio submitted. Speaker labels +$0.02/hr. Price snapshot: 2026-10-17. Sourced preview; confirm the provider quote. Official pricing ↗ · Feature evidence ↗
Azure Speech batch transcriptionMicrosoft Azure	$18.00	Not benchmarked	$0.18base rate	Recorded files Sources & billing notes Standard batch rate, billed per second. Real-time transcription costs $1 per hour. Price snapshot: 2026-10-17. Sourced preview; confirm the provider quote. Official pricing ↗ · Feature evidence ↗

Price not fully listed 1

These match your requirements, but the base price or the speaker-label charge is not published in a form we can calculate.

Soniox STT Async v5SonioxCheck provider ↗

Might fit, not documented 5

The provider does not document at least one of your requirements. That does not mean it is unsupported: test it or ask the provider.

Grok Speech to TextxAICheck provider ↗

GPT TranscribeOpenAICheck provider ↗

GPT-4o TranscribeOpenAICheck provider ↗

GPT-4o mini TranscribeOpenAICheck provider ↗

GPT-4o Transcribe DiarizeOpenAICheck provider ↗

Accuracy is the word error rate (WER) the open Open ASR Leaderboard measured for the exact same model (English average, or FLEURS for the selected language); it is not a test of your audio, and rows without an exact match say so. Costs are API usage estimates, not all-in account bills. Includes documented request minimums and known speaker-label charges when required. Excludes taxes, plan fees, credits, retries, storage and other extras. Incomplete language evidence is kept out of a confirmed shortlist. Test representative audio before integrating.

How accuracy changes with the recording

Word error rate (WER) is the share of words a model gets wrong; lower is better. The ranking above uses the English average. Meetings (AMI), earnings calls (Earnings22) and hour-long recordings are harder, and models do not degrade equally.

Show WER by recording type for 6 benchmarked endpoints

Endpoint	Average	Meetings	Earnings calls	Long-form
Scribe v2ElevenLabs	5.8%	11.9%	9.4%	6.2%
Chirp 3 (Speech-to-Text V2)Google Cloud	6.6%	11.8%	10.6%	9.6%
EnhancedSpeechmatics	6.9%	14.5%	12.3%	7.0%
Whisper Large v3Groq · Model score; the host's decoding settings can differ.	7.4%	15.9%	11.3%	8.7%
Whisper Large v3 TurboCloudflare Workers AI · Model score; the host's decoding settings can differ.	7.8%	16.1%	11.6%	8.7%
Whisper Large v3 TurboGroq · Model score; the host's decoding settings can differ.	7.8%	16.1%	11.6%	8.7%

Only exact model matches get a score; for example AssemblyAI Universal-3.5 Pro, Deepgram Nova-3 and OpenAI's GPT transcribe models are not on the leaderboard in the version we price. A benchmark is not your audio: test names, numbers and accents on your own recordings.

Keep comparing

Other transcription needs

German transcription - Speaker labels - All transcription APIs