Deepgram vs AssemblyAI: Speech-to-Text API Comparison

November 7, 2026

Deepgram and AssemblyAI are both proprietary, hosted speech-to-text APIs that handle recorded files and live audio. They differ in their current flagship models (Deepgram Nova-3, AssemblyAI Universal-3.5 Pro), in how they price streaming and add-on features, and in language coverage. Neither is the accuracy winner for every recording. The practical way to choose is to run both on your own audio and compare errors, speaker labels, latency, and cost.

This guide compares the two providers using the published facts we could verify, the capabilities listed in the Speechall model catalog, and a reproducible test you can run with one API key. Prices and capabilities change. Treat every figure here as a snapshot from November 2026 and confirm it on the linked page before you budget.

The short answer

• Choose by workflow first. Both providers support recorded-file and live transcription, but their streaming models, language lists, and add-ons are different products.
• Compare the models you would actually use. “Deepgram vs AssemblyAI” usually means Nova-3 vs Universal-3.5 Pro for new projects, or Nova-2 vs Universal-2 for cost-sensitive workloads.
• Measure on your own audio. Accents, noise, crosstalk, and domain vocabulary change results more than any general benchmark.

Models compared in the Speechall catalog

The table lists the four models most people compare, with the rates and capabilities shown in the Speechall model catalog. Rates are Speechall credit usage per audio hour. Credit purchase fees are shown separately in your billing quote.

“Languages listed” counts the languages in the Speechall catalog for each model. The provider may support more languages in its own API. AssemblyAI, for example, describes Universal-2 as supporting 99 languages on its pricing page. Check each model page for the exact language codes and the streaming language list.

Pricing from each provider

Buying directly from a provider gives you its own list prices, free allowance, and volume plans. As of November 2026:

Deepgram (pricing page):

• Nova-3 pre-recorded audio costs $0.0043 per minute for monolingual and $0.0052 per minute for multilingual audio on pay-as-you-go.
• Nova-3 streaming costs $0.0077 per minute for monolingual and $0.0092 per minute for multilingual audio.
• New accounts get a free $200 credit with no credit card and no expiration.
• A Growth plan with annual prepaid credits starts at $4,000 per year and lists savings of up to 20%.
• Speaker diarization appears among the add-ons billed on top of transcription.

AssemblyAI (pricing page):

• Universal-3.5 Pro costs $0.21 per hour, and Universal-2 costs $0.15 per hour, for pre-recorded audio.
• Universal-Streaming is listed at $0.15 per hour.
• Speaker diarization is a $0.02 per hour add-on.
• The free tier includes up to 185 hours of pre-recorded and 333 hours of streaming transcription, with no credit card.

Per-minute and per-hour prices are not directly comparable until you convert them. Deepgram’s $0.0043 per minute is $0.258 per hour. Add the features you need, such as diarization, before comparing totals.

Speechall’s catalog rate for a model is not always the same as the provider’s direct list price, and it can be higher. Universal-2 is one example. If you will only ever use one model from one provider, compare your expected monthly volume at both prices. The speech-to-text API cost guide walks through that estimate.

Differences that matter in practice

Flagship models

Deepgram positions Nova-3 as its general-purpose model for both recorded and live audio, with monolingual and multilingual versions. AssemblyAI positions Universal-3.5 Pro as its most accurate model for recorded audio. It describes native code-switching across 18 languages and improved speaker diarization in its Universal-3.5 Pro announcement. We cover the Speechall side of that model in AssemblyAI Universal-3.5 Pro on Speechall.

Streaming

Both providers offer realtime APIs, but they use separate streaming models with their own pricing. AssemblyAI lists dedicated realtime models such as Universal-Streaming. Deepgram streams with Nova-3 and Flux. If live captions or voice agents are your main use case, compare the streaming models and their language lists specifically, not the batch models.

Benchmarks

Both companies publish comparisons. Deepgram has a benchmarks article, and AssemblyAI publishes an AssemblyAI vs Deepgram comparison. Each one is a vendor’s own claim, made with its own datasets and methods. They are useful for understanding what each vendor optimizes for, not as a neutral ranking.

Test both on your own audio

The most reliable comparison is the same recording sent to both models. With Speechall, the two requests differ only in the model query parameter:

# Deepgram Nova-3
curl -X POST "https://api.speechall.com/v1/transcribe?model=deepgram.nova-3&language=en&output_format=json&diarization=true" 
  -H "Authorization: Bearer $SPEECHALL_API_KEY" 
  -H "Content-Type: audio/mpeg" 
  --data-binary @meeting.mp3 -o nova-3.json

# AssemblyAI Universal-3.5 Pro
curl -X POST "https://api.speechall.com/v1/transcribe?model=assemblyai.universal-3-5-pro&language=en&output_format=json&diarization=true" 
  -H "Authorization: Bearer $SPEECHALL_API_KEY" 
  -H "Content-Type: audio/mpeg" 
  --data-binary @meeting.mp3 -o universal-3-5-pro.json

Then compare the two outputs against a transcript you have checked by hand:

1. Calculate word error rate (WER) on a sample of 10 to 30 representative recordings, not one clip.
2. Check speaker attribution on overlapping speech and short turns.
3. Look at names, product terms, and numbers, which matter most to downstream users.
4. Record processing time and the cost per audio hour for each model.

The getting started guide covers API keys and the request format.

When to choose each

Deepgram may fit better when:

• live transcription is central and you want one model family for both batch and streaming;
• you need a language that is in Deepgram’s list but not AssemblyAI’s; or
• your tests show better results on your audio for Nova-3 or Nova-2.

AssemblyAI may fit better when:

• your recordings mix languages within a conversation, which Universal-3.5 Pro is designed for;
• you rely on speaker diarization in recorded meetings or calls and its results test well on your audio; or
• you want to use AssemblyAI’s additional speech understanding features in its own API.

When to integrate directly instead of through Speechall

Speechall is the better fit when you want to compare or switch between models, or use several providers behind one API, key, and bill. Direct integration is often the better fit when:

• you have already chosen one provider and do not expect to switch;
• you need a provider feature Speechall does not expose, such as a specific streaming event or speech understanding add-on;
• the provider’s direct price is lower for your volume, or you have negotiated pricing; or
• your contract or data processing terms must be signed with the provider itself.

The speech API gateway vs direct integration guide covers these tradeoffs in more detail.

Frequently asked questions

Is Deepgram or AssemblyAI more accurate?

It depends on your audio. Both vendors publish benchmarks that favor their own models. Test both on representative recordings and compare word error rate and speaker labels.

Which is cheaper, Deepgram or AssemblyAI?

At direct list prices as of November 2026, Deepgram Nova-3 monolingual pre-recorded audio is $0.258 per hour, AssemblyAI Universal-3.5 Pro is $0.21 per hour, and AssemblyAI Universal-2 is $0.15 per hour. Add-ons such as diarization, streaming, and multilingual audio change the total, so compare the full set of features you need.

Can I use both through one API?

Yes. Speechall offers Deepgram and AssemblyAI models through the same POST /v1/transcribe endpoint. Change the model ID to switch. See the model catalog for the full list.