AssemblyAI Universal-3 Pro vs Deepgram Nova-3: An honest comparison for developers

AssemblyAI's Universal-3.5 Pro and Deepgram's Nova-3 and Flux take different roads to the same destination: turning audio into usable text. This comparison walks through the tradeoffs that actually decide which speech-to-text API fits your build — accuracy, entity handling, latency, languages, developer experience, and pricing — using current benchmarks rather than marketing lines.

Whether you're transcribing medical consultations where drug names have to land perfectly, powering a voice agent that can't afford a garbled credit card number, or processing high-volume archive audio, the right pick depends on what breaks when a word is wrong. We'll keep it factual and specific.

AssemblyAI vs Deepgram at a glance

AssemblyAI and Deepgram both offer async (batch) and real-time speech-to-text APIs. You send audio, you get a transcript back — either after the file finishes processing or streaming word-by-word over a WebSocket. The philosophical split: AssemblyAI's Universal-3.5 Pro leads on accuracy, entity precision, and context features, while Deepgram leans on speed and its Flux model's turn-detection story for voice agents.

As of mid-2026, AssemblyAI's flagship is Universal-3.5 Pro for async and Universal-3.5 Pro Realtime for streaming — the latter is also the speech foundation under the Voice Agent API. Both share the model id universal-3-5-pro. Deepgram fields Nova-3 for general transcription and Flux, which it positions as "conversational speech recognition" for voice agents.

How do accuracy and speed compare?

Accuracy in real-world conditions

Accuracy is usually reported as Word Error Rate (WER) — the share of words the model gets wrong, so lower is better. On the audio you'll actually see in production — phone recordings, background noise, accents, industry jargon — Universal-3.5 Pro leads by a wide margin.

On async multilingual audio with mid-sentence code-switching (think Hinglish or Spanglish support calls), Universal-3.5 Pro posts a normalized WER of 7.69% averaged across five language pairs, versus 12.22% for Deepgram Nova-3 Multilingual. It also beats ElevenLabs Scribe v2 (8.77%), Soniox v5 (10.01%), and OpenAI's GPT-4o Transcribe (44.58%).

For real-time voice-agent audio, the gap is even larger. On the Pipecat open speech-to-text benchmark — real agent conversations — Universal-3.5 Pro Realtime lands a pooled WER of 6.99%, compared to 15.58% for Deepgram Flux.

Entity accuracy is where deals are won or lost

The single most consequential difference isn't overall WER — it's entity accuracy. Those are the details that break downstream workflows when they're wrong: credit card numbers, phone numbers, account IDs, spelled-out names, medical terms. If a voice agent hears "555-0143" instead of "555-0123," the transaction fails.

On real-time agent audio, Universal-3.5 Pro Realtime's overall entity error rate is 15.31% versus Deepgram Flux's 50.50%. Broken out:

This isn't abstract. HotelPlanner switched from Deepgram to AssemblyAI specifically over credit-card and spelled-data accuracy, and reported a 10% lift in checkout conversion after the move. When your agent takes payments or confirms bookings by voice, entity precision is the metric that matters.

Diarization — knowing who said what

Universal-3.5 Pro ships AssemblyAI's most accurate speaker diarization yet, using a joint transcript-plus-speaker model optimized for concatenated minimum-permutation WER (cpWER) rather than the older DER metric. On async audio it averages a cpWER of 30.17 across conditions, versus 37.92 for Deepgram Nova-3 English and 35.26 for ElevenLabs Scribe v2. In realtime, diarization includes an end-of-stream re-clustering revision (about 0.5s after the call ends) for up to 10 speakers.

Speed and latency

Both platforms process batch audio fast enough for most jobs — a 30-minute recording typically finishes in a few minutes on either. For voice agents, though, streaming latency is the number that decides whether a conversation feels natural. Universal-3.5 Pro Realtime's end-of-turn detection reads tonality, pacing, and rhythm and lands around 300ms, with three modes (min_latency, balanced, max_accuracy) so you can tune the tradeoff per use case.

Prompting and customization — the most underrated differentiator

This is where the platforms diverge most. Deepgram offers keyword boosting — a list of terms you want recognized better. It's simple and useful, but it's a list.

Universal-3.5 Pro layers several context mechanisms on top of keyterm prompting:

• Contextual prompting: prime the model with domain context or prior transcript. In an internal healthcare test, feeding prior-visit notes cut missed medical terms by 31%.
• Rolling conversation memory: on by default in realtime, nothing to configure — the model carries earlier turns forward so it interprets later audio in context.
• agent_context: pass your voice agent's own question so the model hears the caller's reply through that lens. Across 20,000 voice-agent files, passing agent context cut WER by 10.2% (fabrications down 18.3%, hallucinations down 17.2%, place entities down 15.5%). One team that paired agent context with prompting drove utterance error rate from 26% to 9% on production audio.
• voice_focus: isolates the primary speaker — near_field for headsets and phones, far_field for rooms, kiosks, and drive-thrus.

The Speech Understanding stack (sentiment analysis, PII redaction, entity detection) plus the LLM Gateway also live on the same platform, so you don't need separate APIs for post-processing.

Which platform fits which team?

The right answer depends on who you are and what you're building. Here's a persona and use-case breakdown.

Medical and clinical transcription

For healthcare workloads, Universal-3.5 Pro pairs its base accuracy with Medical Mode (the medical-v1 domain) and contextual prompting tuned for clinical audio — drug names, dosages, procedures, and clinical terminology. In internal testing, priming the model with prior-visit notes cut missed medical terms by 31%.

On the compliance side: AssemblyAI is considered a business associate under HIPAA and offers a Business Associate Addendum (BAA), which is available to sign without a sales call. That covers workloads involving protected health information when the BAA is in place. For a deeper dive, see the medical transcription solution.

Pricing and total cost

AssemblyAI keeps pricing simple: $0.21/hr for async transcription and $0.45/hr base for realtime, both flat, with no volume commitment and no separate charge for concurrency. Realtime add-ons are transparent — diarization with revision is +$0.12/hr, prompting +$0.05/hr, voice isolation +$0.10/hr. Context features (rolling memory and agent_context) and keyterm prompting are included in the base rate.

Deepgram uses tiered pricing that varies with volume commitments and contract terms, and charges separately for concurrency. That can make it hard to predict a bill as traffic scales. For exact current numbers, check the pricing page.

The headline rate is only part of the story. The hidden cost of lower accuracy shows up as developer time fixing transcripts, support tickets from bad data, failed automations, and lost insight. HotelPlanner's switch — driven by credit-card accuracy — turned into a 10% checkout conversion lift, which dwarfs any per-hour difference.

Developer experience and integration

Both platforms ship REST APIs, Python and Node SDKs, and streaming over WebSockets. AssemblyAI leans hard into developer experience: clear documentation with runnable examples, an interactive playground for testing prompts, robust webhooks, and drop-in plugins for LiveKit and Pipecat. Universal-3.5 Pro Realtime is already live on Retell (high-accuracy mode), LiveKit Inference, and Fireflies. If you're building an agent, one WebSocket handles streaming speech-to-text with context built in, rather than stitching together separate vendors.

Which platform should you choose?

Choose AssemblyAI when accuracy has a business cost

If a wrong word breaks a transaction, a compliance record, or an analytics pipeline, the accuracy and entity gains pay for themselves. Voice agents that take payments, contact-center analytics, medical documentation, and financial services all fall here. Multiple companies — including HotelPlanner — have migrated from Deepgram after benchmarking on their own audio.

When Deepgram might still fit

Deepgram remains reasonable for high-volume, low-stakes transcription where a human will review the output anyway — rough meeting notes, casual voice memos, generic archive content — or if you're deeply integrated and not ready to migrate. Its Flux turn-detection model is purpose-built for agent turn-taking, though its raw accuracy and entity numbers trail Universal-3.5 Pro. Before defaulting to the status quo, test the entity accuracy gap against your actual use case.

Final words

The choice really comes down to a question about your own product: what happens the moment a word is wrong? If the answer is "nothing much," speed and price lead and either platform works. If the answer is "a payment fails, a chart lies, or a clinician misses a dose," entity accuracy becomes the whole game — and that's the metric where Universal-3.5 Pro's lead is widest and most durable. Benchmark both on your audio; the numbers on your data are the only ones that decide it.

Frequently asked questions

AssemblyAI vs Deepgram — which is more accurate?

AssemblyAI's Universal-3.5 Pro is more accurate across the benchmarks that matter for production. On async code-switching audio it posts a 7.69% WER versus Deepgram Nova-3 Multilingual's 12.22%, and on realtime agent audio it hits 6.99% WER versus Deepgram Flux's 15.58%. The gap is largest on entity accuracy — 15.31% entity error rate versus 50.50% for Flux.

Is AssemblyAI or Deepgram better for voice agents?

AssemblyAI is the stronger choice for voice agents on accuracy and entity handling. Universal-3.5 Pro Realtime lands a 6.99% pooled WER on the Pipecat agent benchmark versus Flux's 15.58%, detects end-of-turn around 300ms, and includes agent_context and rolling conversation memory. Deepgram's Flux is built around turn detection but trails on raw accuracy, especially entities like phone numbers and credit cards.

AssemblyAI vs Deepgram pricing?

AssemblyAI charges a flat $0.21/hr for async and $0.45/hr base for realtime, with no volume commitment and unlimited concurrency at no extra charge. Deepgram uses tiered pricing that varies with volume and contract terms, plus separate charges for concurrency. For current figures, see the pricing page.

Deepgram Nova-3 vs Universal-3.5 Pro?

Universal-3.5 Pro is AssemblyAI's flagship async model and the successor to Universal-3 Pro. Against Deepgram Nova-3 it leads on code-switching WER (7.69% vs 12.22% Multilingual) and diarization cpWER (30.17 vs 37.92 for Nova-3 English), and it supports 18 languages with mid-sentence code-switching, contextual prompting, and keyterm prompting where Nova-3 offers keyword boosting.

Which is better for medical or clinical transcription?

AssemblyAI is the stronger fit. Universal-3.5 Pro offers Medical Mode plus contextual prompting that cut missed medical terms by 31% when primed with prior-visit notes in internal testing. AssemblyAI is considered a business associate under HIPAA and offers a Business Associate Addendum (BAA), available to sign without a sales call, which is required before processing protected health information.

How do I migrate from Deepgram to AssemblyAI?

Start by getting a free API key and running Universal-3.5 Pro on a sample of your own audio to compare accuracy and entity results side by side. AssemblyAI provides REST and streaming APIs, Python and Node SDKs, and drop-in plugins for LiveKit and Pipecat, so most integrations are a matter of swapping the transcription endpoint and mapping response fields. See the documentation to map Deepgram parameters to their AssemblyAI equivalents.

‍