Deepgram vs AssemblyAI: Batch Price and Hard-Audio Error

For prerecorded English audio, AssemblyAI Universal-3.5 Pro is the stronger starting choice than Deepgram Nova-3: its listed base rate is lower, and a reproducible batch benchmark measured fewer word errors on both clean speech and difficult earnings calls. That result applies to the tested models and recordings; it does not establish which streaming API will work better in a live conversation.

The choice changes with the workload. Recorded files are priced by audio duration, while a live application must also account for its streaming model, open sessions, simultaneous calls and required languages. Keeping those questions separate makes the published prices and error rates useful.

What a prerecorded hour costs

AssemblyAI’s pay-as-you-go schedule lists prerecorded Universal-3.5 Pro at $0.21 per audio hour and Universal-2 at $0.15. Universal-3.5 Pro supports 18 languages and Universal-2 supports 99; prerecorded speaker diarization adds $0.02 per hour on either model, while keyterms prompting adds $0.05 on Universal-3.5 Pro. For live transcription, the same schedule lists Universal-Streaming at $0.15 per hour and Universal-3.6 Pro Realtime at $0.45.

Deepgram’s pay-as-you-go pricing lists prerecorded Nova-3 Monolingual at $0.0043 per minute and Multilingual at $0.0052, equivalent to $0.258 and $0.312 per hour. Its promotional Nova-3 streaming rates are $0.0048 and $0.0058 per minute, equivalent to $0.288 and $0.348 per hour. Prerecorded speaker diarization is included; streaming diarization is an add-on. Deepgram describes the Nova family as supporting more than 45 languages, with availability depending on the model.

For a hypothetical 1,000 hours of prerecorded audio, base transcription comes to $210 on Universal-3.5 Pro, $150 on Universal-2, $258 on Nova-3 Monolingual or $312 on Nova-3 Multilingual. Add-ons, multiple audio channels and contracted discounts could change an invoice. Universal-2 has the lowest listed batch rate in this group, but the accuracy measurements below belong to Universal-3.5 Pro; assigning them to the cheaper model would make the comparison misleading.

Where the error gap comes from

In the Verli speech-to-text benchmark, AssemblyAI Universal-3.5 Pro recorded 1.55% word error rate on LibriSpeech test-clean and 10.43% on Earnings-22, compared with 2.50% and 16.13% for Deepgram Nova-3. The clean set contains read audiobook speech; Earnings-22 contains accented financial earnings calls. Both engines transcribed the same files, and the benchmark provides transcripts and scoring code.

The difference between these models is 0.95 percentage points on clean speech and 5.70 points on the harder calls. Word error rate counts substitutions, deletions and insertions against a reference transcript, so it describes how often words differ, not how costly each difference is. A wrong customer name or amount may matter more to an application than a missed filler word, even when each adds to the score.

The wider gap on Earnings-22 makes Universal-3.5 Pro a credible first candidate for difficult prerecorded English audio. Earnings calls are still a particular kind of difficult recording: they do not establish either provider’s error rate on a room microphone, overlapping meeting speakers or an accented customer calling from a noisy street. The benchmark also evaluates batch transcription, so its ranking cannot be carried over to the vendors’ live models.

Live audio changes the bill

For a hypothetical 1,000 billable streaming hours at the listed base rates, Universal-Streaming costs $150, Universal-3.6 Pro Realtime $450, Nova-3 Monolingual $288 and Nova-3 Multilingual $348. These are different models with different features, not four prices for the same accuracy. Deepgram labels its Nova-3 streaming rates promotional, which matters when estimating a recurring cost beyond the current offer.

AssemblyAI’s Universal-Streaming billing guidance states that an open WebSocket session is billed for its full duration, including idle time, and that simultaneous sessions accrue charges separately. A voice application that leaves a connection open between utterances can therefore have more billable session time than spoken-audio time. The relevant cost estimate uses expected connection duration and concurrent sessions, rather than call volume alone.

Live model selection also needs a different performance measure. A voice agent may depend on how quickly it receives usable words, when a final transcript arrives and how well the model handles interruptions. None of those outcomes follows from the prerecorded word-error figures, even if the same provider name appears on both products.

Limits and languages can decide the fit

Deepgram’s pay-as-you-go rate limits allow up to 50 concurrent Nova-3 prerecorded requests and 150 concurrent Nova-3 streaming requests per project in the listed regions. Streaming requests with speaker diarization have a lower service limit: 50 in North America and 25 in the other listed regions. A burst of calls or a backlog of recordings may reach those limits before the per-hour price becomes the binding constraint.

Language support needs to be checked at the model level. AssemblyAI’s broader Universal-2 language range does not give Universal-3.5 Pro the same coverage, and Deepgram prices its monolingual and multilingual Nova-3 options separately. For mixed-language calls, the useful comparison is between models that actually accept the required languages and handle switching within the recordings the application receives.

A small test for the audio that matters

A team choosing between these APIs can build a labeled set from recordings it is permitted to use, with a human-checked reference transcript for each clip. Keep clean and difficult conditions separate; a large group of easy clips can conceal a recurring failure in the calls that matter most. The set should include:

Clear, close-microphone speech to establish an ordinary baseline.
Accents, speaking speeds and language switches that reflect actual users.
Overlapping speakers, room microphones and background noise from the intended meeting or call setting.
Names, amounts and domain terms whose errors would affect the product’s output.

Run each clip through the exact batch or streaming models under consideration, using comparable language and prompting settings where possible. Record word error rate separately for each condition, note failed or missing transcriptions, and inspect business-critical terms rather than relying on one aggregate score. If speaker labels or live responses matter, assess those outputs directly alongside the transcript.

For prerecorded English work that fits Universal-3.5 Pro’s language range, its lower listed rate and lower error on both published test sets make it the initial candidate. A different model can still earn the choice through better results on the team’s recordings, required language coverage or live behavior. The deciding cost figure should come from the same workload used to judge the output.