Deepgram vs AssemblyAI: Speech-to-Text APIs Compared
November 18, 2026

Deepgram vs AssemblyAI used to be a simple accuracy-versus-speed argument. In late 2026 it's a different question: do you want one vendor to run the whole voice pipeline, or the best transcript you can get to feed your own? Deepgram sells speech-to-text, text-to-speech and a bundled Voice Agent API, and it will run inside your own cloud. AssemblyAI has just renamed and replaced its flagship models, and it now points voice agent builders to open frameworks instead of its own agent API.
The short version, as of November 2026:
- Pick Deepgram if you're building a real-time voice agent and want speech-to-text, the LLM step and text-to-speech behind one WebSocket, or if you need to self-host both batch and streaming transcription.
- Pick AssemblyAI if recorded audio is your main job (meetings, calls, podcasts, medical notes) and you want prompting, speaker labels and LLM post-processing from the same API key.
- Test both on your own audio. Accuracy and speed depend on accents, audio quality and vocabulary. Neither vendor's benchmark page will tell you how your recordings behave.
Model names in this category change every few months, and most comparison posts are already stale. Every name below was checked on each vendor's own docs on November 18, 2026.
Deepgram vs AssemblyAI: quick comparison
| Deepgram | AssemblyAI | |
|---|---|---|
| Current speech-to-text models | Flux STT (voice agents), Nova-3 (general) | AssemblyAI-STT-3.7 (pre-recorded), STT-3.7-Realtime and STT-3.7-Realtime-Agent (streaming), Universal-2 |
| Languages, per docs | Flux STT Multilingual: 10. Nova-3: 55 in multilingual mode plus a longer monolingual list | STT-3.7: 32. Universal-2: 99 |
| Billing unit | Per minute, billed by the second | Per hour |
| Starting offer | $200 credit that does not expire | Hours of pre-recorded and streaming transcription at no charge |
| Self-hosting | Yes, containers in your cloud or data centre, for Enterprise | Streaming only, with an upfront commercial commitment |
| Bundled voice agent API | Yes, Voice Agent API | Being retired; LiveKit and Pipecat reference builds instead |
| LLM features on transcripts | Audio Intelligence | LLM Gateway (LeMUR is gone) |
| Text-to-speech | Flux TTS, Aura-2 | AssemblyAI-TTS-1 |
Current models, as of November 2026
Deepgram: Flux STT and Nova-3
Deepgram's models and languages overview lists two current families for new work.
Flux STT is the streaming model built for voice agents. Deepgram describes it as conversational speech recognition with turn detection built into the model, so it decides when the caller has finished speaking instead of leaving that to a separate component. It comes as an English model and a multilingual model covering 10 languages.
Nova-3 is the general-purpose model for both batch and streaming. It has a multilingual mode for audio where speakers switch languages, a long list of single-language options, and medical and pharma variants. The list is still growing: Deepgram's changelog shows 36 more Nova-3 languages added on November 16, 2026.
Nova-2 is still listed, and Deepgram recommends it for languages Nova-3 doesn't cover yet and for filler word detection. Nova, Enhanced and Base sit under "Legacy models". Deepgram's pricing page says new projects should start with Nova-3 for general transcription or Flux for real-time voice agents.
AssemblyAI: the STT-3.7 family replaced Universal-3.5 Pro
AssemblyAI changed its lineup the day before this post was written. Its models page now lists:
- AssemblyAI-STT-3.7, the flagship for pre-recorded audio. 32 languages, native code switching, and prompting that lets you describe the audio in plain language.
- Universal-2, the lower-cost pre-recorded model, covering 99 languages. You can list both models in one request so anything STT-3.7 doesn't support falls through to Universal-2.
- AssemblyAI-STT-3.7-Realtime-Agent, the streaming model for voice agents.
- AssemblyAI-STT-3.7-Realtime, the streaming model for notetakers and ambient transcription, with realtime speaker labels.
- Universal-3.6 Pro Realtime, the previous streaming flagship, still supported.
AssemblyAI's model versions and deprecations page marks universal-3-5-pro and universal-3-pro as deprecated on November 17, 2026, with assemblyai-stt-3-7 as the replacement. For pre-recorded audio the retirement date is still to be announced. The older streaming models carry a tentative retirement date of February 1, 2027.
If a tutorial tells you to use Universal-3.5 Pro or Universal-3 Pro, it predates this change. Deprecated models keep working for now, and retired ones are routed to their replacement, but new code should name the current model IDs.
Languages and accuracy
The two vendors count languages differently, so read the lists and not the totals.
AssemblyAI's top model supports 32 languages, and its broader model reaches 99. Deepgram's Nova-3 multilingual mode lists 55 languages, and its monolingual list is longer and was extended this month. Flux STT, Deepgram's agent model, covers 10.
What that means in practice:
- Mainstream languages (English, Spanish, French, German, Hindi, Japanese, Portuguese): both vendors cover them with their best models.
- Less common languages: check the exact language code on each vendor's list. A language may be supported only by an older or lower-tier model.
- Mixed-language audio: both offer code switching, on Nova-3 Multilingual and Flux STT Multilingual for Deepgram, and on AssemblyAI-STT-3.7, STT-3.7-Realtime-Agent and Universal-2 for AssemblyAI.
We're deliberately not quoting word error rates. Both companies publish benchmarks that show their own model winning, and results swing with audio quality, accents and jargon. Take 20 to 30 minutes of your real audio, run it through both, and count the errors that matter to you: names, numbers, product terms.
Pricing: per minute vs per hour
Both are usage-based with no minimum, and both let you start without a credit card. The units differ, so convert before you compare.
Deepgram bills per minute of audio, by the second, with a lower rate card on its prepaid Growth plan and custom pricing for Enterprise. The Deepgram pricing page puts monolingual streaming at $0.29 an hour for Nova-3 and $0.39 for Flux STT on pay-as-you-go, says pre-recorded runs lower, and labels some current rates as promotional. Features such as redaction, keyterm prompting and speaker diarization are per-minute add-ons.
AssemblyAI prices per hour. The AssemblyAI pricing page lists AssemblyAI-STT-3.7 at $0.21 an hour and Universal-2 at $0.15 an hour for pre-recorded audio, and $0.30 an hour for STT-3.7-Realtime and $0.45 an hour for STT-3.7-Realtime-Agent for streaming. Speaker diarization, prompting, medical mode and PII redaction are per-hour add-ons.
| Deepgram | AssemblyAI | |
|---|---|---|
| Unit | Per minute, per-second billing | Per hour |
| General streaming model | Nova-3, $0.29/hr | STT-3.7-Realtime, $0.30/hr |
| Voice agent streaming model | Flux STT, $0.39/hr | STT-3.7-Realtime-Agent, $0.45/hr |
| Add-ons | Per minute | Per hour |
| Volume option | Growth, prepaid annual credits | Custom terms through sales |
Rates are from each pricing page on November 18, 2026 and both pages change often.
The base rates are close enough that add-ons decide the bill. Work out the price of the exact feature set you need (diarization plus redaction plus keyterms, say) on both pages. A transcript with every option turned on can cost well over the headline rate on either side.
Voice agents and LLM features
This is the biggest practical difference today.
Deepgram has a bundled agent product. Its Voice Agent API handles listening, thinking and speaking over a single WebSocket, with function calling and telephony guides for Twilio, Plivo, Genesys and Amazon Connect. It's priced per minute of connection time, with lower rates if you bring your own LLM or text-to-speech provider.
AssemblyAI is stepping back from its own agent API. AssemblyAI's model reference for AI assistants states that its Voice Agent API is being retired and shouldn't be used for new builds. The recommended path is its LiveKit or Pipecat reference implementations, which combine the Realtime-Agent speech model, LLM Gateway and its new text-to-speech API. You get the parts from one vendor, and an open framework does the orchestration. That page gives no retirement date, so check with AssemblyAI if you already run on the old API.
If you want a fully managed phone agent on top of either speech model, that's a different layer of the stack. Our Vapi vs Retell AI comparison covers it.
LeMUR is gone; LLM Gateway replaced it
Plenty of older posts still list LeMUR as a reason to choose AssemblyAI. The AssemblyAI changelog announced that LeMUR would be deprecated on March 31, 2026 and stop working after that date, and the old LeMUR doc URLs now redirect to LLM Gateway.
LLM Gateway is an OpenAI-compatible endpoint that reaches Claude, GPT and Gemini models with one key. You use it to summarise, translate, or ask questions about a transcript. It works well for call summaries and meeting notes without adding a second vendor.
Deepgram's equivalent is Audio Intelligence (summaries, topics, intents, sentiment), which is narrower than a general LLM endpoint. For open-ended questions about a transcript on Deepgram, you'd bring your own LLM.
Self-hosting, regions and compliance
Deepgram self-hosts more. Its self-hosted docs describe running Deepgram in containers on your own cloud instances or bare metal, or through Amazon SageMaker. In a typical deployment, no audio or transcripts go to Deepgram, only licence checks and usage metadata. The pricing page says self-hosting is available on Enterprise and needs NVIDIA GPUs.
AssemblyAI documents self-hosting for streaming only. Its self-hosted streaming page describes a Docker Compose stack with GPU-backed containers, where audio and transcripts stay in your network. It also states the offering requires an upfront commercial commitment of $50,000. We found no documented self-hosted option for AssemblyAI's pre-recorded API.
On regions, both run US and EU endpoints. AssemblyAI's docs list AWS regions in the United States and the EU for pre-recorded audio. Deepgram offers an EU endpoint, and its changelog shows an India endpoint reaching general availability in September 2026.
On compliance, as each vendor states it:
- Deepgram's pricing page lists SOC 2 Type 1 and Type 2, HIPAA with a BAA for Enterprise customers, GDPR, CCPA and PCI, plus FIPS 140-3 images for self-hosted deployments.
- AssemblyAI's docs state SOC 2 Type 1 and Type 2 certification and refer to a Business Associate Addendum for customers who need one. They also say files from customers under that addendum, on EU servers, or opted out are not used for model training.
Ask each vendor for its reports, and read the data-retention and model-training defaults yourself before you send real customer audio.
AssemblyAI vs Deepgram: which one should you pick?
| Your situation | Better fit |
|---|---|
| Real-time voice agent, one vendor for the full pipeline | Deepgram |
| Meetings, calls or podcasts transcribed after the fact | AssemblyAI |
| Summaries and Q&A on transcripts without another vendor | AssemblyAI, with LLM Gateway |
| Batch transcription that must run on your own servers | Deepgram |
| A less common language | Whichever lists your exact language code |
| You already build on LiveKit or Pipecat | Either. Both plug in as the speech-to-text step |
If neither is right, we keep roundups of Deepgram alternatives and AssemblyAI alternatives. And if you'd rather run an open model yourself, the Whisper alternatives post covers that route.
Where Taku fits
Deepgram and AssemblyAI are APIs. You still have to build the product around them. If you'd rather run a tool someone already built than set one up yourself, that's where Taku comes in. Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.
FAQ
Is Deepgram or AssemblyAI more accurate?
It depends on your audio. Both vendors publish benchmarks that favour their own models. Run the same sample of your real recordings through both and compare the errors on names, numbers and domain terms.
What is AssemblyAI's best model right now?
As of November 2026, AssemblyAI's docs name AssemblyAI-STT-3.7 as the flagship for pre-recorded audio and AssemblyAI-STT-3.7-Realtime-Agent as the recommended streaming model for voice agents. Universal-3.5 Pro was deprecated on November 17, 2026.
What is Deepgram's best model right now?
As of November 2026, Deepgram recommends Flux STT for real-time voice agents and Nova-3 for general transcription. Nova-2 remains for languages Nova-3 doesn't yet support.
Is AssemblyAI's LeMUR still available?
No. AssemblyAI's changelog set March 31, 2026 as the date LeMUR would stop working. Its replacement is LLM Gateway, which uses a chat-completions style API.
Can I self-host Deepgram or AssemblyAI?
Deepgram offers self-hosted containers for Enterprise customers. AssemblyAI documents a self-hosted option for streaming transcription only, with an upfront commercial commitment.
Which is cheaper, Deepgram or AssemblyAI?
Base rates are close once you convert per-minute to per-hour. The difference comes from add-ons such as diarization, redaction and keyterm prompting, and from volume discounts. Price your exact feature set on both pricing pages.