Speech-to-Text API Comparison: Price per Hour, Diarization and Data Use
Data checked:
A speech-to-text API is bought on four numbers: the price per minute or hour, what speaker labels add, which languages a model actually covers, and what happens to your audio by default. That last one varies more than price: some providers only use requests for training if you ask them to, others put you in an improvement program unless you opt out. All rates below are list prices from vendor pages checked in October 2026; per-hour conversions are ours, and volume tiers and regions change the real bill.
| Tool | Languages | Speaker labels | Trains on your data? |
|---|---|---|---|
| AssemblyAI API, File transcription, Live transcription | ENJAZH 99 languages | Paid plans | Yes, unless you opt out |
| Deepgram API, File transcription, Live transcription | ENJAZH | Yes | Yes, unless you opt out |
| Speechmatics API, File transcription, Live transcription | ENJAZH 56 languages | Yes | Only if you opt in |
| OpenAI Speech to Text API API, File transcription, Live transcription | ENJAZH | Yes | Only if you opt in |
| Google Cloud Speech-to-Text API, File transcription, Live transcription | ENJAZH | Yes | Only if you opt in |
| Amazon Transcribe API, File transcription, Live transcription | ENJAZH | Yes | Yes, unless you opt out |
| ElevenLabs Speech to Text (Scribe) API, File transcription, Live transcription | ENJAZH 99 languages | Yes | Yes, unless you opt out |
Prices
- AssemblyAI: Universal-2 costs $0.15 per audio hour, usage-billed; speaker diarization adds $0.02 per hour.
- Speechmatics: Melia 1 prerecorded transcription is $0.12 per hour.
- Google Cloud Speech-to-Text: V2 Standard Dynamic Batch is $0.003 per minute (about $0.18 an hour); regular V2 Standard starts at $0.016 per minute (about $0.96). The famous 60 free minutes a month apply to V1 standard recognition, not V2.
- OpenAI Speech to Text API: gpt-4o-mini-transcribe is $0.003 per minute (about $0.18 an hour) and gpt-transcribe is $0.0045 (about $0.27); the guide marks the cheaper model deprecated, with shutdown on 26 February 2027.
- Deepgram: Nova-3 monolingual is $0.0043 per minute (about $0.26 an hour); multilingual is $0.0052 (about $0.31).
- ElevenLabs Speech to Text (Scribe): Starter is a $6 monthly subscription that includes 1 hour 31 minutes a month in the UI; extra API hours cost $0.22. The free tier lists 12 minutes a month.
- Amazon Transcribe: usage-billed per audio second, with 60 free minutes a month for 12 months. The entry-level unit price was not readable in the pricing table we retrieved, so check the AWS page for your region before comparing.
- Rev sells a separate speech-to-text API through Rev AI; its pricing is not in our data.
Speaker labels
Diarization is included at Deepgram, Speechmatics, Amazon Transcribe and ElevenLabs Speech to Text (Scribe), and available on selected models and regions at Google Cloud Speech-to-Text (including Japanese chirp_3). At AssemblyAI it is a paid add-on ($0.02 per hour). At OpenAI Speech to Text API speaker labels require gpt-4o-transcribe-diarize, a model the guide marks deprecated until 26 February 2027.
Upload and request limits
OpenAI limits file uploads to 25 MB, which is the tightest constraint on this list for long audio; ElevenLabs lists 3 GB per file. No per-file duration limit was verified for the others.
What happens to your audio
- Used for training only if you opt in: OpenAI Speech to Text API (API data is not used unless you explicitly opt in to share it) and Google Cloud Speech-to-Text (audio is not logged by default; opting into data logging permits training use).
- Used for training unless you opt out: Deepgram (requests join the Model Improvement Program by default; set mip_opt_out=true), AssemblyAI (paid accounts can opt out in Data Controls; a signed BAA opts out automatically), Amazon Transcribe (voice inputs are stored and used by default; an AWS Organizations opt-out policy disables it) and ElevenLabs Speech to Text (Scribe) (can be disabled in account settings).
- Speechmatics does not train on your data unless you enable it in the portal workspace, and opting in grants a perpetual, irrevocable licence, so read that screen before flipping it.
Language coverage
Counts and coverage are per model, not per vendor: AssemblyAI lists 99 languages for Universal-2 (Chinese not individually verified), ElevenLabs lists 99 for Scribe v2, Deepgram lists English, Japanese and Chinese for Nova-3, Amazon lists English, Japanese, Mandarin and Cantonese with feature differences, Google lists English and Japanese (Chinese not verified for the models we checked), and Speechmatics lists 56 for Melia 1 without a language table we could verify individually. OpenAI does not publish a count on the pages we checked.
Self-hosting instead
If none of the data-use defaults fit, whisper.cpp runs Whisper locally with a C/C++ interface, free and open source; language coverage depends on the model file you load. See also the overview in best transcription software or compare every tool in the directory.