Deepgram vs AssemblyAI: transcription cost and accuracy

AssemblyAI Universal-3.5 Pro costs $0.21 per audio hour versus $0.258 for Deepgram Nova-3 and has the lower AA-WER (3.0% vs 5.2%); Nova-3 processes audio 5.5× faster (548.1× vs 100.3× real time).

Updated · 5 sources · Methodology

DimensionWinnerMarginEvidence
Price / audio hourAssemblyAI$0.048$0.258 / $0.2123
AA-WER v2AssemblyAI2.2 pp5.2% / 3.0%1
Speed factorDeepgram5.5×548.1× / 100.3×1

Performance

ProviderModelAA-WER v2 (%)Speed factor
Deepgram1Nova-35.2%548.1×
AssemblyAI1Universal-3.5 Pro3.0%100.3×

Prices and request cost

ProviderPrice / audio hour100 h
Deepgram2$0.258$25.80
AssemblyAI3$0.21$21.00
Cost example: 100 hours of mono pre-recorded audio
Deepgram$25.80
AssemblyAI$21.00

Capabilities and serving conditions

DimensionDeepgramAssemblyAI
Multilingual price and language scopeCode-switching audio uses Nova-3 Multilingual at $0.0052/min ($0.312/h), above the compared $0.0043/min mono rate; Nova-3 Multilingual supports 55 languages.25Universal-3.5 Pro’s listed $0.21/h rate already includes native code-switching across its 18 languages, with no separate multilingual rate.34

When each option fits

Deepgram

  • Use Nova-3 for large mono recordings when completing the queue sooner matters more than the last increment of transcription accuracy.1
  • Its 5.2% AA-WER is higher than Universal-3.5 Pro’s 3.0%; low-error archival transcripts favor AssemblyAI.1

AssemblyAI

  • Choose Universal-3.5 Pro when recorded conversations mix its supported languages and contextual prompting is useful.4
  • Its 100.3× processing factor trails Nova-3’s 548.1×, so the lower price does not also buy the fastest queue completion.1

Methodology and sources

AA-WER v2 uses AA-AgentTalk (50%), cleaned VoxPopuli (25%) and cleaned Earnings22 (25%). Speed factor is AA’s 7-day median of audio seconds processed per second on 10-minute audio, not realtime latency. Prices are base mono pre-recorded rates, excluding add-ons and extra channels.

Audio charges are estimates at listed rates, excluding tax and subscription credits. Novita AI publishes this speech API comparison using public sources.

  1. Artificial Analysis · AA-WER v2 non-streaming
  2. Deepgram · official API pricing
  3. AssemblyAI · official API pricing
  4. AssemblyAI · Universal-3.5 Pro languages
  5. Deepgram · Nova-3 multilingual code-switching

Build with Novita Model APIs

Novita AI doesn’t offer speech-to-text. Pair your transcription provider with Novita LLM APIs to summarize and analyze transcripts, or Fish Audio text-to-speech for voice output.

Start BuildingExplore Models