Speech-to-text API prices per hour, compared (2026)

As of 2026-09-27, cloud speech-to-text costs about $0.04 to $0.40 per hour of audio for batch transcription: Groq Whisper turbo $0.04, AssemblyAI Universal-2 $0.15, ElevenLabs Scribe v2 $0.22, Deepgram Nova-3 about $0.26, OpenAI GPT Transcribe $0.27. Streaming usually costs more, up to $1.02 an hour for OpenAI’s live model. For one person dictating 4 hours a month, that is $0.16 to $1.60.

Price per hour table, 11 providers

List prices from each provider’s own page, checked 2026-09-27. Per-minute prices are multiplied by 60. “4 h/month” uses the batch price.

Provider	Model	Batch $/hour	Streaming $/hour	4 h/month	Source
Groq	Whisper large-v3 turbo	$0.04	n/a	$0.16	Groq
Soniox	V5	$0.10	$0.12	$0.40	Soniox
xAI	Grok speech-to-text	$0.10	$0.20	$0.40	xAI
AssemblyAI	Universal-2 / Universal-3.5 Pro	$0.15 / $0.21	$0.45 (Universal-3.5 Pro only)	$0.60 / $0.84	AssemblyAI
Mistral	Voxtral Mini Transcribe V2	$0.18	$0.36	$0.72	Mistral
ElevenLabs	Scribe v2	$0.22	$0.39	$0.88	ElevenLabs
Deepgram	Nova-3	$0.26	$0.29 promo ($0.46 regular)	$1.04	Deepgram
OpenAI	GPT Transcribe / GPT Live Transcribe	$0.27	$1.02	$1.08	OpenAI
Google	Gemini 3.5 Transcribe	about $0.30	about $0.54	about $1.20	Google
Speechmatics	Enhanced	$0.40	$0.43	$1.60	Speechmatics
Cartesia	Ink 2	Credit plans, no flat rate	Credit plans	n/a	Cartesia

Google prices Gemini transcription in audio tokens, so its per-hour figure is an estimate. Cartesia sells credits on monthly plans (its $5 Pro plan includes a little over 9 hours) instead of a per-minute price.

Batch vs streaming prices

Batch is cheaper everywhere it is offered separately. Deepgram’s gap is small at the promo streaming rate ($0.0043 against $0.0048 a minute); the regular (non-promo) streaming rate is $0.46 an hour, more than double the batch price. OpenAI’s is large: GPT Transcribe is $0.0045 a minute, GPT Live Transcribe $0.017, almost four times as much.

For dictation, streaming matters less than it sounds. A 20-second dictation sent as a finished file comes back quickly on most of these APIs. Streaming starts work while you are still talking, which helps with long dictations. If cost matters more than the last fraction of a second, use batch.

What 4 hours of dictation a month costs

Four hours of speech a month is a heavy personal user: about 12 minutes of talking every workday. At list prices:

• Groq Whisper turbo: $0.16.
• ElevenLabs Scribe v2: $0.88.
• OpenAI GPT Transcribe: $1.08 batch, $4.08 streaming.
• Speechmatics Enhanced: $1.60.

Compare that with $12 to $15 a month for a dictation subscription, and a local model on your Mac at $0. The provider fee is not where dictation gets expensive.

If you add a cleanup pass (punctuation, fillers, spoken corrections) through a small language model, that costs under a cent an hour on the model Nota uses by default. The speech-to-text step is most of the bill.

Hidden costs: minimums, rounding, file limits

• File size limits. OpenAI’s transcription endpoint takes files up to 25 MB, about 13 minutes of typical audio. Longer files fail unless you split them.
• Rounding. Deepgram bills per second with no rounding up. Check the others’ terms if you send many short clips; a per-request minimum adds up across a hundred 5-second dictations a day.
• Different models, different names. OpenAI still lists older models (whisper-1 and gpt-4o-transcribe at $0.006 a minute, gpt-4o-mini-transcribe at $0.003). Pick by the model id you actually call.
• Language support. The cheapest model is only cheap if it handles your language. Cartesia Ink 2 is English only; Voxtral covers 13 languages; Whisper-based and the larger models cover most.
• Plans vs pay-as-you-go. Some list prices are for pay-as-you-go; volume plans and credits change the rate.
• Prices move. Some of these rates differ from what the same pages showed a week earlier. Recheck before you build a budget on them.

Using these APIs from a Mac dictation app

A dictation app that takes your own key calls the provider directly, so you pay the list price above and nothing on top.

In Nota, 11 cloud speech providers are built in, plus any OpenAI-compatible endpoint. Paste a key in the provider’s panel and press Verify. The Models table shows a cost per hour for each cloud model before you pick it, and Home shows what each provider has cost you this month, split into speech-to-text and cleanup. Your provider’s own bill is the final word.

Local models (Parakeet, Whisper) cost $0 and need no key. Setup is in the cloud models docs.