Speech-to-Text API Pricing: Cost per Hour of Audio

Machine transcription is cheap and human transcription is not. Across the 34 data points ¹ in this dataset, drawn from 15 sources ², the median list price for recorded audio is 0.21 US dollars per hour ³. Real-time streaming has a median of 0.46 US dollars per hour ⁴. The two human services we could price have a median of 83.7 US dollars per hour ⁵.

Every figure is a vendor's own list price, converted to US dollars per hour of audio and read in October 2026. Negotiated, promotional and sales-call prices are left out. For background on what these services do and how accurate they are, see our overview of transcription.

Recorded audio costs cents per hour, and the spread is wide

The batch rows for recorded audio run from 0.04 US dollars per hour at the cheapest, Groq's turbo Whisper model ⁶, to a high end set by an AI subscription product at 0.02 US dollars per hour ⁷. The median sits between them at 0.21 US dollars per hour ³.

Most dedicated speech-to-text vendors cluster close to that median:

• AssemblyAI lists its flagship model at 0.21 US dollars per hour ⁸ and its older model at 0.15 ⁹.
• ElevenLabs lists Scribe at 0.22 US dollars per hour ¹⁰, and the same price applies to its medical variant ¹¹.
• Soniox lists file transcription at 0.10 US dollars per hour ¹².
• Rev's AI service lists English transcription at 0.20 US dollars per hour ¹³.

The big cloud platforms and OpenAI sit a little higher or lower depending on the model. Amazon Transcribe batch comes to 0.006 US dollars per hour ¹⁴, the same as OpenAI's whisper-1 at 0.006 ¹⁵. OpenAI's smaller transcription model and Mistral's Voxtral both list 0.003 US dollars per hour ^{16 17}, and Google's dynamic batch rate matches at 0.003 ¹⁸.

Streaming costs more than recorded audio at most providers

Real-time transcription has a median of 0.46 US dollars per hour ⁴, higher than the batch median of 0.21 ³. The reason is practical: a live stream needs a model that is always ready and answers in a fraction of a second, while a file can be queued and run on cheap capacity.

Same-provider pairs show the premium:

All values are US dollars per hour of audio. The exception is AssemblyAI, whose English and multilingual streaming models are listed at 0.15 US dollars per hour ²³, below its flagship recorded-audio price. Soniox has the lowest streaming price in the data at 0.12 US dollars per hour ²⁴. OpenAI's live transcription model has the highest at 0.017 US dollars per hour ²⁵.

Human transcription costs far more than any machine service

The human rows are few but far from the machine prices. Scribie's homepage starts at 0.80 US dollars per hour ²⁶. Rev's human service, with an accuracy guarantee and a stated turnaround, lists 119.40 US dollars per hour ²⁷. Even the cheaper of the two is far above the most expensive machine row.

The gap buys something specific: a person checks the text, and the vendor guarantees a stated accuracy. Machine prices buy a draft that you check yourself. Human rows also carry add-ons that machines mostly do not. Scribie's page, for example, charges extra for priority processing and for noisy or accented audio, and those surcharges are not in the rows.

The same Whisper model is priced very differently by different hosts

Open-weight Whisper models are sold by several hosts, and the prices disagree:

• Groq lists whisper-large-v3 at 0.111 US dollars per hour ²⁹.
• Fireworks AI lists its large model at 0.0015 ³⁰ and the turbo variant at 0.0009 ³¹.
• Rev AI lists Whisper Large for English at 0.30 ³².
• OpenAI lists whisper-1 at 0.006 ¹⁵.

The model is similar, so the difference is the service around it: hardware, minimum billing per request, rate limits and support. Groq bills a minimum length per request and Rev AI rounds up to the nearest second with its own minimum, which matters if you send many short clips.

Tiers and commitments move the price only modestly

Where a provider publishes a second tier, the discount is real but not dramatic. Deepgram's Growth tier, which needs pre-paid credits, lists its monolingual batch model at 0.216 US dollars per hour ³³, against 0.258 on pay as you go ²¹. Its multilingual streaming model drops from 0.552 to 0.468 US dollars per hour ^{34 35}.

Deepgram also advertises lower promotional prices. We recorded only its regular prices, so a live bill may be below these rows.

Where the numbers disagree or cannot be compared

• Different models, different quality. A row is a price, not a quality grade. A cheaper model may make more errors, and you pay for the corrections in your own time. Price per hour says nothing about accuracy on your accents, vocabulary or noise.
• Different billing units. Vendors quote per minute, per hour or per token estimate. We multiplied per-minute prices by 60. OpenAI's figures are labeled as estimated costs and do not say which audio length they assume.
• Narrow examples. Amazon's rows come from a worked example for one US region and one volume tier, not a full price list. Prices vary by region.
• Possibly stale figures. Fireworks' rows come from a launch post, and its current pricing page does not list audio. Mistral's figure is a launch price. Treat these as indications.
• Not like for like. The AssemblyAI synchronous API is limited to short clips and is priced above its file-upload model at 0.30 US dollars per hour ³⁶. The GoTranscript row is an AI product on a subscription plan, and its pay-as-you-go euro price has no row.

The dataset also covers a small number of providers. It lists no enterprise discounts, which can be large, because those need a sales call. Add-ons such as speaker labels, translation and custom vocabulary are priced separately and are not included.

What this means if you need text from audio

• Recorded meetings, interviews and podcasts: use a batch API or a free in-browser tool. At a median of 0.21 US dollars per hour ³, price rarely decides the choice. Test two or three engines on your own audio. Our audio to text page transcribes files in the browser at no charge.
• Live captions or voice agents: expect to pay more. Compare streaming rows only, and check the latency and language support behind each price.
• Legal, medical or publication work that must be right: a human service costs far more, but it comes with an accuracy guarantee. A common approach is machine first, then a human pass only on the passages that matter.
• Dictating your own writing: you do not buy transcription by the hour at all. Dictation apps, including Yap, which we make, run speech recognition on your device, and the built-in tools on your computer or phone are often enough. See our overview of transcription for how these options differ.

Check the live price page before you commit. These rows are a snapshot from October 2026, and vendors change list prices often.