How accurate is Deepgram?

Deepgram is accurate, but not the most accurate speech-to-text API on public English tests. On the independent Artificial Analysis index, Nova-3 scored a 5.2 percent word error rate (WER) ¹. WER is the share of words a system gets wrong, so lower is better and 5.2 percent means roughly 5 wrong words in every 100. On the same index AssemblyAI's Universal-3.5 Pro scored 3.0 percent ².

Treat any single WER as a guide, not a promise. Your result depends on your audio, and the sources below disagree by several points.

On an independent index, Nova-3 trails the newest models

Artificial Analysis runs paid APIs on three test datasets with one method, which makes it the fairest like-for-like comparison we found. Its figures:

The same Nova-3 model scored 4.8 through Telnyx and 5.2 through Deepgram's own API. Settings and test dates differ between listings, so a half-point gap should not decide anything. A gap of two points, as between Nova-3 and the leaders, is more meaningful.

Deepgram's own and rivals' benchmarks disagree

At launch, Deepgram reported a 5.26 percent median WER for Nova-3 on batch audio and 6.84 percent on streaming in its own tests ⁷. AssemblyAI's comparison page, on its own test, gives Nova-3 an average of 7.9 percent against 5.93 percent for its Universal-3.5 Pro ⁸. Neither vendor publishes its test audio, so neither can be checked, and each favors its own product. Our question on why accuracy claims differ explains how the choice of test set moves the numbers.

Older evidence is weaker, and live audio trades accuracy for speed

A 2024 academic study of university lecture recordings measured Deepgram at an average 8.3 percent WER on English, against 2.9 percent for Whisper large-v2 ⁹. That study predates Nova-3, so it describes earlier models, not the current one.

On live audio, Deepgram's Flux model was the fastest service in Artificial Analysis's 2026 streaming benchmark, returning a final transcript 0.020 seconds after speech ended, at a WER of 7.36 percent ¹⁰. Its first partial transcript had a 7.4 percent WER, against 3.6 percent for ElevenLabs' realtime model ¹¹. For a voice agent that must reply quickly, that tradeoff can be right. For captions people will read, it may not be.

Deepgram also does not appear in the English results of Hugging Face's Open ASR Leaderboard that we read on 20 October 2026, so there is no direct comparison there.

It depends: what changes your own error rate

• Audio quality. Noise, distance from the microphone and overlapping speakers raise WER for every system.
• Vocabulary. Names and jargon are the usual failures. Deepgram's Keyterm Prompting, its form of custom vocabulary, accepts up to 100 terms per request ¹².
• Language. Public scores are mostly English. Test your language directly.
• Scoring rules. Whether punctuation, capitals and filler words count changes WER a lot. See what counts as a good word error rate.

What to do

1. Do not choose on a leaderboard alone. Collect 30 to 60 minutes of your real audio and write correct reference transcripts.
2. Run it through Nova-3, add your key terms and run it again, then score both with WER.
3. Run the same files through AssemblyAI and one other service. The $200 credit covers the test ¹³.
4. If live speed matters more than the last point of accuracy, weigh Flux separately.

Continue with Deepgram vs AssemblyAI, how much Deepgram costs, the guide to Deepgram and the overview of speech recognition. For the best-scoring options overall, see the question on the most accurate speech-to-text API.