2 models compared on the same benchmark — accuracy, latency, price, language coverage and capabilities. Nova-2 leads on raw accuracy; Nova-2 is fastest; Nova-3 covers the most languages; Nova-2 is cheapest.
Verdict · who wins what
Most accurate
Nova-2
31.5% WER
Deepgram
Fastest
Nova-2
0.1× realtime
Deepgram
Most languages
Nova-3
47 languages
Deepgram
Cheapest
Nova-2
$0.0058/min
Deepgram
Comparing
2 models · same benchmark
1
Nova-2
Deepgram
Deepgram's Nova-2 speech recognition
2
Nova-3
Deepgram
Deepgram's flagship model — 53% lower WER vs competitors, code-switching support
30-day benchmark average
overall WER vs price
Nova-2
Nova-3
field · 26 models
lower-left is better
WER · lower is better
WER · English
p50 → p99
Streaming benchmark averages
Rolling 30 days
billed per second
Feature support
Languages · formats · regions
general
legal
medical
noisy
technical
uk
us
Latency p99
2.8s
1.6s
1.02×
—
Flicker
84.7%
—
Cadence
0.8/s
—
$34.92
$46.08
1,000 hours
$349.20
$460.80
Yes
Auto-detect language
Yes
Yes
Live / streaming
Yes
Yes
Custom vocabulary
Yes
Yes
Compliance
Is Nova-2 or Nova-3 more accurate?
Nova-2 is more accurate, with a 31.5% word error rate versus 36.8% for Nova-3, on our standardized benchmark.
Which is cheaper, Nova-2 or Nova-3?
Nova-2 is cheaper at $0.0058/min versus $0.0077/min for Nova-3.
Should you choose Nova-2 or Nova-3?
Choose Nova-2 for accuracy, speed, and lower cost; choose Nova-3 for language coverage.