Deepgram Nova-3 Is the Fastest Batch Speech-to-Text Model on Artificial Analysis, but Not the Most Accurate

Artificial Analysis ranks Deepgram Nova-3 first on batch speed, with a median speed factor of 601.2 as of 16 October 2026. The same leaderboard puts its word error rate at 5.2%, behind Scribe v2 at 2.2% and Universal-3 Pro at 3.1%. For voice agents, speed is the stronger case when inputs are short and predictable, and accuracy matters more when one wrong word changes the outcome.

Disclosure: This post covers speech-to-text evaluation and voice AI, a market that overlaps with Humanbit's work in data labelling, voice data and AI evaluation.

How fast is Deepgram Nova-3 in batch transcription?

On the non-streaming leaderboard, Nova-3 posts a median speed factor of 601.2, captured 16 October 2026. The leaderboard times 10-minute recordings, so that factor implies a file of that length finishes in about one second.

The batch timing on the methodology page includes the network round trip, so the figure is not pure model compute. Each published speed figure is a seven-day median, which means the number shifts as new trials arrive. Any speed claim needs a capture date.

Deepgram describes speed in different terms. Its launch post says Nova-3 latency is comparable to Nova-2 and up to 40 times faster than competing models that include diarization. That is the company's own claim.

How accurate is Nova-3 compared with other speech-to-text APIs?

Nova-3 is not among the most accurate models on Artificial Analysis's index. As of 16 October 2026, its 5.2% AA-WER trails Scribe v2 (2.2%), Universal-3 Pro (3.1%) and GPT Transcribe (3.3%).

Source: Artificial Analysis non-streaming leaderboard, captured 16 October 2026. Prices are Artificial Analysis's normalised estimates per 1,000 minutes, not list prices.

Nova-3 costs more than Scribe v2 and Universal-3 Pro on this page, and less than GPT Transcribe. The same model also varies by host. Nova-3 served by Telnyx posts 4.8% and a speed factor of 265.8, at 7.40 USD per 1,000 minutes. Deepgram's endpoint is about 2.3 times faster on the page, while the Telnyx copy costs about 72% more and is slightly more accurate. Host and configuration choices can move results as much as the model name does, a pattern also seen in harness-sensitive coding agent scores.

The AA-WER index is a weighted average. Half of its weight comes from AA-AgentTalk, a proprietary held-out set of 469 samples. VoxPopuli-Cleaned-AA (628 samples) and Earnings22-Cleaned-AA (six long recordings) carry 25% each, per the methodology page. Outside readers cannot inspect half the score, and a quarter of it rests on six recordings.

How does Nova-3 perform on streaming latency?

On Artificial Analysis's streaming benchmark, Nova-3 is among the fastest models tested. Its Realtime variant reaches a first partial transcript in 0.057 seconds at 6.69% WER, and a final transcript at 0.06 seconds at 6.7% WER, according to the streaming article.

Deepgram's Flux model is faster still, at 0.020 seconds, but its WER is 7.36%. ElevenLabs's Scribe v2 Realtime reaches 3.65% on first partials, at 0.13 seconds. Deepgram's benchmark page says Nova-3 returns transcripts in under 300 ms, which it presents as the bar for real-time voice agents.

Accuracy leaders on the streaming board come from other vendors. Cartesia Ink-2 with semantic endpoints posts 3.59% WER at 0.21 seconds, and ElevenLabs Scribe v2 Realtime posts 3.64% at 0.14 seconds, per the same streaming article. No single model leads on every dataset in that article.

Coval's Pulse page, which also carries a Nova 3 row, was last measured on 16 October 2026 over a 30-day window. It ranks Nova 3 fifth of 29 on time to final segment, at 85 ms. The same page ranks it 11th of 27 on time to first token, at 1,433 ms, and 20th of 31 on WER, at 5.3%. Speed rankings depend on the metric, so one leaderboard position does not settle the question.

Where do Deepgram's own numbers and independent tests disagree?

Deepgram reports a median batch WER of 5.26% and a median streaming WER of 6.84% for Nova-3. Both figures come from its own test set. The independent boards place Nova-3 close to the batch figure, but they use different audio.

The company's dataset holds 2,703 audio files from nine domains, amounting to 81.69 hours. Against a next-best competitor at 10%, Deepgram claims a 47.4% batch improvement. On streaming, it claims 54.2% against a competitor at 14.92%. The launch post gives the streaming figure as 54.3% in its summary, so the company's own numbers do not match.

Independent boards now show 5.2% for Nova-3 on Artificial Analysis and 5.3% on Coval. Those values sit close to Deepgram's batch figure, but they do not confirm its dataset result, because each board uses its own audio, scoring and normalisation. Relative percentages also depend on the competitor chosen, so they say little on their own. This vendor-versus-independent gap follows the pattern in lab-reported Terminal-Bench scores.

When does speed beat accuracy for a voice agent?

Speed wins when inputs are short and predictable and response delay is the main complaint. Accuracy wins when one wrong word changes the outcome, such as an account number, a dose or a booking time.

The accuracy gap is measurable. On Artificial Analysis's batch index, 5.2% against 2.2% means about three more wrong words per 100, or roughly 2.4 times as many errors. On streaming, Nova-3 Realtime sits at 6.7% at the final transcript, while Scribe v2 Realtime sits at 3.6%, per the streaming article.

• Yes or no answers and menu picks. Artificial Analysis says its first-partial metric suits quick answers such as yes or no. That is where Nova-3's latency matters most.
• Names, numbers, addresses and product codes. Test accuracy on your own audio first. Deepgram says keyterm prompting accepts up to 100 terms at request time, a company claim aimed at this kind of vocabulary.
• Mixed call flows. Measure both. Coval places Nova 3 near the top on time to final segment but in the lower half on accuracy, so one ranking will not decide the choice.

To check the trade on your own audio, follow these steps:

1. Pull a representative sample of real recordings, including the names and codes your agent must capture.
2. Send the same audio to Nova-3 and to a rival at matching endpoint settings, so the harness does not decide the result.
3. Normalise both transcripts with one scorer, so punctuation and number formatting do not count as word errors.
4. Score first partial and final transcripts separately, because a fast partial can be less accurate than the final.

How are these benchmarks run, and who runs them?

Streaming timing depends on endpointing. Where a model accepts a forced endpoint signal, Artificial Analysis sends that signal when its voice activity detector fires. Models without that option use a fallback rule, and the methodology page reports partial and final WER separately for that reason.

Artificial Analysis tests API endpoints directly and publishes its methodology, including dataset weights and sample counts. Coval runs daily tests on hosted endpoints, and it also sells evaluation services, so it has a commercial stake in the category. Its Pulse page invites readers to book a demo for testing their own production calls.

Artificial Analysis counts network delay in every streaming latency figure, so the results reflect what end users experience. Coval's Nova 3 row covers 26,022 samples over 30 days. The AA batch index covers about 8 hours of audio across three datasets. Neither source reports a contamination check for Nova-3's training data, so that question remains open.

Deepgram's figures come from its own suite, and its pages promote its products. Treat them as claims until an outside lab reproduces them.

What's still unclear?

Four points remain open. The speed factor depends on the snapshot. Half the AA index is proprietary. Deepgram's launch post gives two streaming reductions, 54.3% and 54.2%. No benchmark here covers Nova-3 on a specific production audio mix.

The weakest point of the case is comparability. Speed and accuracy come from different boards, hosts and datasets, so the 601.2 figure and the 5.2% figure each describe one snapshot. Your own audio can move the result more than any leaderboard cited here. Artificial Analysis's streaming benchmark is also recent, having launched on 1 June 2026.