Deepgram vs AssemblyAI vs Rev AI: choosing a speech-to-text API

Pre-recorded audio: compare the base rate and the add-ons

Deepgram prices per minute, AssemblyAI and Rev AI mostly per hour. On pay-as-you-go list prices checked on November 7, 2026, Deepgram's Nova-3 Monolingual costs $0.0043 a minute (about $0.26 an hour) and Nova-3 Multilingual $0.0052 a minute. AssemblyAI's Universal-3.5 Pro costs $0.21 an hour and Universal-2 $0.15. Rev AI's Reverb costs $0.20 an hour for English and $0.30 for other languages.

Add-ons can change the order. Deepgram includes speaker diarization and smart formatting on pre-recorded audio and charges $0.0020 a minute for redaction. AssemblyAI adds $0.02 an hour for diarization, $0.08 for PII text redaction and $0.05 for keyterms prompting on Universal-3.5 Pro. Rev AI's pricing page lists insights such as language identification and summarization separately. It does not list a separate price for diarization.

Rev AI also sells human transcription through the same API, at $1.99 a minute. Deepgram's Growth plan starts at $4,000 a year in prepaid credits for up to 20% off, and Rev AI offers volume-based Enterprise pricing.

Streaming: the billing clock matters as much as the rate

AssemblyAI's Universal-Streaming is $0.15 an hour, and Universal-3.6 Pro Realtime is $0.45. AssemblyAI bills streaming on how long the WebSocket stays open, not on audio sent, so idle time counts. Unclosed sessions close automatically after 3 hours and are billed for all 3. Deepgram's Nova-3 Monolingual and Flux STT English stream at $0.0077 a minute, which Deepgram states as $0.46 an hour. Deepgram bills per second of processed audio, with no minimum and no rounding up.

Rev AI bills each stream on the larger of stream duration and audio duration, rounded up to the second, with a 15-second minimum. Its pricing page does not show a streaming-specific rate. Check the rate that applies to streaming with Rev AI before you estimate costs.

Concurrency defaults differ. Deepgram allows 150 concurrent streaming connections on pay-as-you-go. AssemblyAI's pay-as-you-go plan starts at 100 new streaming sessions per minute and raises that limit automatically as usage approaches it. Rev AI's default streaming concurrency limit is 10, adjustable by support.

Languages: batch and streaming coverage differ

Deepgram says its Nova models support 60+ languages. Its automatic language detection covers 35 languages for pre-recorded audio. AssemblyAI's coverage depends on the model: Universal-2 lists 99 languages and Universal-3.5 Pro lists 18 for pre-recorded audio. For streaming, Universal-3.6 Pro Realtime supports 32 languages, Universal-Streaming is English only and Universal-Streaming Multilingual covers six.

Rev AI lists 57+ languages for asynchronous transcription, but its streaming API accepts 9: English, French, German, Italian, Japanese, Korean, Mandarin, Portuguese and Spanish. Match your languages to the specific model: AssemblyAI’s six-language Universal-Streaming Multilingual covers fewer languages, while its 32-language Universal-3.6 Pro Realtime covers more.

Data retention, model training and HIPAA

By default, Deepgram retains request audio and transcripts to improve its models. Adding mip_opt_out=true to a request means its audio and transcripts are kept only while the request is processed; request metadata and usage logs are still recorded. Deepgram signs BAAs for Enterprise customers and offers EU, Australia and India endpoints. Its regional data guarantees require attention to MIP opt-out: data retained for model improvement can be transferred outside the selected region.

AssemblyAI lets paid accounts opt out of model training, set a retention TTL and sign a BAA from the dashboard, at no extra cost. Free accounts cannot use these controls. Signing a BAA opts the account out of training permanently. With no TTL or BAA, uploaded audio starts deletion at 24 hours and transcripts at 30 days. Accounts opted out of training get zero retention of streaming audio and transcripts.

Rev AI keeps asynchronous jobs for up to 30 days. You can shorten this per account or per job, or delete jobs through the API. HIPAA processing requires a signed BAA, an updated MSA and a separate account; under it, human transcription and media URLs are not supported. Rev AI's EU deployment in Frankfurt is arranged through support.

Which one to try first

Consider Deepgram for real-time voice products that need many languages, per-second billing on audio rather than connection time, or regional endpoints. Opt out of model improvement if you need zero retention. Consider AssemblyAI for low-cost streaming, clearly itemized add-ons or self-serve HIPAA and training controls. Consider Rev AI for low-cost English batch transcription, a human transcription option or forced alignment. Run your own audio through each before deciding, because these list prices do not measure accuracy.

This comparison reviews official documentation and prices checked November 7, 2026. No audio tests, latency measurements or accuracy benchmark were performed.

Sources & method

Official documentation was checked on 2026-11-07. Selection advice is our assessment. Interfaces and requirements may change; check the linked source for the version you intend to use.

• Deepgram pricing (streaming and pre-recorded tabs, FAQ)
• Deepgram languages
• Your Data at Deepgram: retention, MIP opt-out and regional endpoints
• Deepgram Model Improvement Partnership Program
• Deepgram data privacy compliance (HIPAA, GDPR)
• AssemblyAI pricing (models, add-ons, guardrails, FAQ)
• AssemblyAI billing and pricing
• AssemblyAI data retention and model training
• AssemblyAI Data Controls: opt-out, TTL and BAA
• Rev AI pricing
• Rev AI supported languages
• Rev AI streaming request options and languages
• Rev AI streaming billing and credits
• Rev AI FAQ
• Rev AI security and privacy (job deletion)
• Rev AI HIPAA compliance
• Rev AI global deployments
• Deepgram current diarization model and response contracts

How SOTA reviews projects · Suggest a correction · Download the evaluation checklist