Speech-to-text API pricing, compared per audio hour

Speech-to-text vendors quote the same service in three different units. Some price per second of audio, some per minute, and others per hour. Google Cloud, Amazon Transcribe and Azure quote per second or per minute, Deepgram and AssemblyAI quote per hour, and Soniox publishes per-second rates that look small until multiplied by 3,600. The tables on this page convert everything to dollars per hour of audio.

The bigger split is between batch and streaming. Batch transcribes a file you already have, and streaming returns text while the audio is still being spoken. Streaming is usually priced higher, and several vendors sell only one of the two. 8 of the 33 models here publish both lanes. The columns stay separate, and the monthly estimate uses whichever lane you select.

Two further complications affect the real bill. Some models are priced in tokens instead of time: OpenAI's gpt-4o-transcribe family charges for the audio tokens it reads and the text tokens it writes, and Google's Gemini transcription models do the same, so the hourly figure is an estimate. 5 rows here are estimates and carry a tilde. Extras are also billed separately at many vendors: speaker labels, language identification, redaction and custom vocabulary are often add-ons, and some services set a minimum billed duration per request.

Each model's published rate is converted to dollars per hour of audio. A per-second price is multiplied by 3,600, a per-minute price by 60, and a per-hour price is used unchanged. For token-priced models the page uses the per-minute figure the vendor itself quotes and multiplies by 60, instead of guessing at token counts.

• Batch and streaming are never averaged or merged. A model with only one lane shows a dash in the other column and is hidden when you select the lane it lacks.
• Speaker labels are added as a separate per-hour add-on where the vendor charges for them, and a model without labels is hidden when you select Needed.
• Promotional prices are used as published and flagged, since they can end without notice. One row is flagged as promotional now.

All prices are list prices in US dollars. Volume tiers, committed-use contracts and free credits are not applied, so a large account can pay materially less than the table shows.

Consider a team transcribing 1,000 hours of recorded calls a month. On OpenAI Whisper Large v3 Turbo the batch bill is $30.60. On the median batch model (Mistral Voxtral Mini Transcribe 3) it is $210. On Google Cloud Speech-to-Text V2 (Chirp 3) it is $960. The difference between the cheapest and priciest batch price is about 31 times.

Live transcription of the same 1,000 hours costs $90 at NVIDIA Nemotron 3 ASR Streaming 0.6B and $1,020 at OpenAI GPT Live Transcribe. Streaming bills the time the stream is open, including silence and hold music, so a call center that keeps streams open for the whole call pays for the whole call, not only for speech.

Preparing the audio changes the bill as much as the choice of vendor. Dropping silent stretches, splitting a stereo call into one channel per speaker only when the vendor charges per channel, and avoiding re-uploads of the same file all reduce billed hours. Where a vendor offers both lanes, transcribe recordings in batch even if the same model also streams, because the batch rate is usually lower.

The monthly estimate is simply audio hours times the hourly rate, plus any speaker-label add-on. Enter your own hours above, or use the 100, 1,000 and 10,000 hour presets to see how the ranking holds at different scales. It rarely changes at the extremes, but minimum durations start to matter once you process very many short clips.

Accuracy is not in the table. Word error rate depends heavily on the audio: accents, background noise, domain vocabulary and recording quality can move results more than the choice of vendor. For published comparisons on shared test sets, see the Hugging Face Open ASR Leaderboard (opens in new tab), then run your own sample before committing.

Volume discounts are also left out. Google Cloud's Chirp 3 list price steps down through volume tiers to about $0.24 per hour, and several vendors negotiate custom rates above a monthly commitment. Deepgram Nova-3 is priced promotionally for now, so check the vendor page before relying on it. Speaker labels are a billed add-on at 2 models and bundled at 7. OpenAI Whisper Large v3 (10 s), OpenAI Whisper Large v3 Turbo (10 s), Rev.ai Reverb (15 s) bill a minimum duration per request.

Language coverage, custom vocabulary, punctuation and formatting, redaction of sensitive data and latency to the first word are also outside the table. They vary widely between vendors and between languages within the same vendor, and a model that is cheapest for English may not support the language you need at all. Treat the price ranking as a shortlist, then test two or three candidates on a sample of your real recordings.

Open-weight Whisper models are hosted by several providers at very different prices. The cheapest open-weight batch row here is OpenAI Whisper Large v3 Turbo at $0.031 per audio hour. You can also run Whisper yourself and pay for GPU time instead of audio hours, which only wins at steady high volume.

Prices come from vendor pricing pages and from the LiteLLM community price list, refreshed daily. Rows are labeled Official, LiteLLM, Reseller or Third-party in the source type column, each links to its source, and the Verified only filter hides third-party and reseller prices. 14 providers are covered. For the full process, see the methodology page.

Not listed: Fireworks speech-to-text (no published price), Scaleway speech-to-text (only a zero placeholder price is listed), Cohere speech-to-text (no published price), Microsoft MAI-Transcribe (no vendor-published price found) and Gladia Solaria (priced in EUR on plan-dependent tiers, so it cannot be put on one USD basis). A model is added once a price we can verify is published.