Speech-to-text pricing looks simple: a price per minute of audio. In practice, providers quote per minute, per hour or per second, split batch from streaming, and sell diarization and redaction as add-ons. This page puts the main APIs on one scale, using prices we fetched from each official page on 14 and 15 November 2026 (each source below gives its date).
All prices are list, pay-as-you-go, in US dollars, for batch (pre-recorded) audio unless noted. Volume discounts and contracts are not included.
How much do speech-to-text APIs cost per minute and per 1,000 hours?
| Provider and model | Per minute | Per hour | Per 1,000 hours |
|---|---|---|---|
| Groq, Whisper Large v3 Turbo | about USD 0.00067 | USD 0.04 | USD 40 |
| DeepSeek ASR, limited-time price (50% off list) | about USD 0.00097 | USD 0.058 | USD 58 |
| Soniox, async (billed per token; “about” per hour) | about USD 0.00167 | about USD 0.10 | about USD 100 |
| Groq, Whisper Large v3 | about USD 0.00185 | USD 0.111 | USD 111 |
| DeepSeek ASR, list price | about USD 0.00193 | USD 0.116 | USD 116 |
| BroadNet Transcribe (14 languages) | $0.002 | $0.12 | $120.00 |
| AssemblyAI, Universal-2 | USD 0.0025 | USD 0.15 | USD 150 |
| OpenAI, gpt-4o-mini-transcribe (deprecated, shuts down 2027-02-26) | USD 0.003 | USD 0.18 | USD 180 |
| Google, V2 dynamic batch | USD 0.003 | USD 0.18 | USD 180 |
| AssemblyAI, Universal-3.5 Pro | USD 0.0035 | USD 0.21 | USD 210 |
| Speechmatics, Melia 1 multilingual (56 languages) | USD 0.004 | USD 0.24 | USD 240 |
| BroadNet Transcribe Arabic | $0.0042 | $0.25 | $252.00 |
| Deepgram, Nova-3 monolingual | USD 0.0043 | USD 0.258 | USD 258 |
| OpenAI, gpt-transcribe | USD 0.0045 | USD 0.27 | USD 270 |
| Deepgram, Nova-3 multilingual | USD 0.0052 | USD 0.312 | USD 312 |
| OpenAI, whisper-1 / gpt-4o-transcribe (deprecated, shut down 2027-02-26) | USD 0.006 | USD 0.36 | USD 360 |
| AWS Transcribe, batch (US East) | USD 0.006 | USD 0.36 | USD 360 |
| Speechmatics, Standard | USD 0.0075 | USD 0.45 | USD 450 |
| Speechmatics, Enhanced | USD 0.0125 | USD 0.75 | USD 750 |
| Google, V2 standard (first 500,000 min) | USD 0.016 | USD 0.96 | USD 960 |
Per-hour and per-1,000-hour figures are our arithmetic from the published unit price (per minute x 60, then x 1,000). Groq, DeepSeek, AssemblyAI and Speechmatics publish per-hour prices; we divided by 60 for the per-minute column. DeepSeek’s pricing page lists its speech-to-text API at USD 0.116 an hour with a limited-time 50% discount and gives no end date, so budget on the list price. Speechmatics rates are its list pre-recorded prices; its pricing page can also show them with a 33% model-training discount and up to 25% volume discount applied, which brings Melia 1 down to about USD 0.12 an hour. The 33% discount is opt-in, and by opting in you let Speechmatics use your audio and transcripts to improve its models (its pricing FAQ, checked 2026-11-15).
Soniox, DeepSeek ASR and Groq’s Whisper endpoints all cost less per hour than both BroadNet models. Soniox’s pricing page, checked 2026-11-14, bills per token (USD 1.50 per 1M input audio tokens and USD 3.50 per 1M output text tokens for async) and says this is “about” USD 0.10 an hour for file uploads and USD 0.12 an hour for real-time streaming. Its language list includes Arabic among more than 60 languages. It uses its own job-based API rather than the OpenAI /v1/audio/transcriptions shape, so switching to it means new client code.
One change to plan for: OpenAI’s deprecations page, checked 2026-11-15, lists whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize as deprecated (notice given August 26, 2026), shutting down on February 26, 2027, with gpt-transcribe or gpt-live-transcribe as the replacement. Its pricing page marks the same four models deprecated. If you price an OpenAI pipeline today, price it on gpt-transcribe at USD 0.0045 a minute. That is more per hour than both BroadNet models, but OpenAI also sells live transcription (gpt-live-transcribe, USD 0.017 a minute), which BroadNet does not. This page compares prices only; if you are moving off whisper-1, the Whisper API alternative guide covers how to test a replacement on your own audio and what breaks in the switch. To run your own volumes against these rates, use the speech-to-text cost calculator.
For the full picture, our open speech-to-text API price index lists batch and streaming prices, billing increments and Arabic support for more than 20 products, each checked against the provider’s own pricing page (CSV and JSON).
What does the base price leave out?
The per-minute number is rarely the whole bill. Check these before you compare:
| Item | What the pages say |
|---|---|
| Speaker diarization | Deepgram: included for pre-recorded, USD 0.0020/min extra for streaming. AWS: included in the standard price. AssemblyAI: +USD 0.02/hr. OpenAI: separate gpt-4o-transcribe-diarize model at USD 0.006/min, deprecated and shutting down 2027-02-26. |
| PII redaction | Deepgram: USD 0.0020/min. AWS: USD 0.0024/min (tier 1). AssemblyAI: text redaction +USD 0.08/hr on Universal-2 and +USD 0.12/hr on Universal-3.5 Pro; audio redaction +USD 0.05/hr. |
| Streaming | AWS: USD 0.01/min vs USD 0.006 batch. Deepgram Nova-3 monolingual: USD 0.0048 streaming (a promotional price valid until 2027-02-13) vs USD 0.0043 pre-recorded. Speechmatics Melia 1: USD 0.40/hr real-time vs USD 0.24/hr pre-recorded. |
| Billing minimum | Groq bills at least 10 seconds per request. AWS bills per second with no minimum. Google rounds each request up to the next second. |
| Volume tiers | Google V2 standard drops to USD 0.01/min after 500,000 minutes a month, and lower above 1 million. AWS standard batch and streaming are now flat rates; its volume tiers apply to add-ons such as redaction. |
For a call centre that needs diarization and redaction, Deepgram Nova-3 pre-recorded becomes USD 0.0063 per minute with redaction, and AWS becomes USD 0.0084. A cheap base price with expensive add-ons can lose to a higher base price with them included.
What do 10,000 voice notes a month cost to transcribe?
Say a support inbox receives 10,000 WhatsApp voice notes a month, 30 seconds each. That is 5,000 minutes.
- At USD 0.006 a minute (AWS batch, or OpenAI’s deprecated whisper-1): USD 30.
- At USD 0.0045 (OpenAI gpt-transcribe): USD 22.50.
- At USD 0.0043 (Deepgram Nova-3): USD 21.50.
- On BroadNet Transcribe Arabic: $21.00.
Billing minimums matter here. Groq’s 10-second minimum does not change the bill for these 30-second notes, but it bills a two-second “ok” note as 10 seconds. With a 15-second minimum, that two-second note costs the same as a 15-second one. For the full voice-note workflow, see voice notes to text for business.
What matters besides price?
A cheaper model that gets 1 word in 5 wrong costs more in human correction than it saves. Before switching, measure word error rate on 30 to 50 of your own recordings per language. Our Whisper alternative guide shows how to run that test in an afternoon, and the Arabic speech-to-text buyer’s guide covers dialect-by-dialect scoring.
Other non-price questions:
- Languages. BroadNet Transcribe covers 14 languages (full list); OpenAI says Whisper (the deprecated
whisper-1) supports 98 with varying accuracy. - File limits. OpenAI caps uploads at 25 MB and BroadNet at 25 MiB, so split long calls.
- Retention. Audio is often more sensitive than text. Ask whether audio and transcripts are stored; see LLM API data retention compared.
Where BroadNet Transcribe fits
BroadNet Transcribe costs $0.002 a minute ($0.12 an hour) across 14 languages: English, Spanish, French, German, Italian, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese and Arabic. BroadNet Transcribe Arabic costs $0.0042 a minute and is built for Arabic dialects; BroadNet reports it #2 of 37 (its September 2026 figure; the live board listed it #4 of 39 on 15 November 2026) on the Open Universal Arabic ASR Leaderboard. Both use the OpenAI-compatible /v1/audio/transcriptions endpoint, accept WAV or MP3 up to 25 MiB, and keep no audio or transcripts (zero data retention). The $5 starting credit covers about 1,190 minutes of Arabic audio; create a key to test your own files.
Where it does not fit: BroadNet does not list streaming transcription, built-in diarization or redaction add-ons. If you need live captions or speaker labels from the API itself, Deepgram, AssemblyAI and AWS offer both. OpenAI offers live transcription, but its only speaker-labelling model is deprecated. Compare the two BroadNet models side by side on the comparison page, or see all prices on pricing.
