DeepgramAudio matchupSpeech to text
Nova-3vs.Flux
Which model is better?
Nova-3 is slightly ahead in independent tests, but the lead is not certain.
Independent testsCoval, Cekura
Nova-30 of 2
Flux0 of 2
Popularityout of 100
Nova-3ahead51
Flux37
Weights and access
Nova-3proprietary
Fluxproprietary
Released
Nova-3n/a
Fluxn/a
Pick Nova-3 for
leaning Covalleaning Cekuramore popular
Pick Flux for
No clear edge
Shared tests
Independent evaluators and leaderboards only, the newest shared snapshot per evaluator. With 95 % ranges a lead counts when they do not overlap; for hit rates, when it holds in a 95 % test. Without any uncertainty figure at most “slightly ahead”.
Coval
Nova-3 slightly aheadNova-3 is clearly ahead in 0 of 23 subtests and slightly in 13, Flux in 0 clearly and slightly in 5. 5 are level.
Coval RTF maximum
Nova-32.89
Flux5.29
lower is better, as of November 5, 2026
Coval AudioToFinal maximum
Nova-329.38 s
Flux17.25 s
lower is better, as of November 5, 2026
Coval Time to Final Segment (TTFS) maximum
Nova-319,227 ms
Flux13,993 ms
lower is better, as of November 5, 2026
Show all 23 subtests
| Subtest | Nova-3 | Flux | Verdict |
|---|---|---|---|
| Coval RTF maximum (lower is better) | 2.89 | 5.29 | Nova-3 slightly ahead |
| Coval AudioToFinal maximum (lower is better) | 29.38 s | 17.25 s | Flux slightly ahead |
| Coval Time to Final Segment (TTFS) maximum (lower is better) | 19,227 ms | 13,993 ms | Flux slightly ahead |
| Coval Time to First Token (TTFT) mean (lower is better) | 1,440 ms | 1,095 ms | Flux slightly ahead |
| Time to Final Segment by dataset: WildASR Accents (lower is better) | 82 ms | 103 ms | Nova-3 slightly ahead |
| Time to Final Segment by dataset: WildASR Clipping (lower is better) | 84 ms | 104 ms | Nova-3 slightly ahead |
| Time to Final Segment by dataset: WildASR Phone codec (lower is better) | 84 ms | 102 ms | Nova-3 slightly ahead |
| Coval Time to First Token (TTFT) maximum (lower is better) | 5,457 ms | 4,616 ms | Flux slightly ahead |
| Coval Time to Final Segment (TTFS) mean (lower is better) | 86 ms | 102 ms | Nova-3 slightly ahead |
| Time to Final Segment by dataset: WildASR Noise gaps (lower is better) | 86 ms | 101 ms | Nova-3 slightly ahead |
| Coval Time to First Token (TTFT) p50 (lower is better) | 1,000 ms | 1,172 ms | Nova-3 slightly ahead |
| Time to Final Segment by dataset: PipeCat Production (lower is better) | 83 ms | 96 ms | Nova-3 slightly ahead |
| Time to Final Segment by dataset: WildASR Clean (lower is better) | 87 ms | 100 ms | Nova-3 slightly ahead |
| Time to Final Segment by dataset: WildASR Far-field (lower is better) | 89 ms | 102 ms | Nova-3 slightly ahead |
| Coval Word Error Rate (WER) per-sample mean (lower is better) | 5.6 % | 6.4 % | Nova-3 slightly ahead |
| Coval Time to Final Segment (TTFS) p50 (lower is better) | 91 ms | 99 ms | Nova-3 slightly ahead |
| Coval pooled corpus Word Error Rate (WER) (lower is better) | 5.2 % | 5.7 % | Nova-3 slightly ahead |
| Coval Word Error Rate (WER) per-sample maximum (lower is better) | 100.0 % | 92.3 % | Flux slightly ahead |
| Coval RTF mean (lower is better) | 1.01 | 1.02 | level |
| Coval AudioToFinal mean (lower is better) | 8.02 s | 8.05 s | level |
| Coval AudioToFinal p50 (lower is better) | 8.39 s | 8.42 s | level |
| Coval RTF p50 (lower is better) | 1.01 | 1.01 | level |
| Coval Word Error Rate (WER) per-sample p50 (lower is better) | 3.1 % | 3.1 % | level |
Source · latest November 5, 2026
Cekura
Nova-3 slightly aheadFinal-text delay (TTFS), P50
Nova-3103 ms
Flux108 ms
lower is better, as of October 23, 2026
Source · latest October 23, 2026
Vendor figures
Figures from vendor reports that cover both models. They do not count toward the verdict.
Show 49 vendor figures
| Figure | Source | Nova-3 | Flux |
|---|---|---|---|
| AssemblyAI Table 4 medical entity error rate (MER), dictation (lower is better) | AssemblyAI Benchmarks | 17.4 % | 18.3 % |
| AssemblyAI Table 4 medical entity error rate (MER), mean 3 sets (lower is better) | AssemblyAI Benchmarks | 13.2 % | 14.3 % |
| AssemblyAI Table 4 medical entity error rate (MER), primary care (PriMock57) (lower is better) | AssemblyAI Benchmarks | 19.0 % | 20.5 % |
| AssemblyAI Table 4 medical entity error rate (MER), respiratory (lower is better) | AssemblyAI Benchmarks | 3.3 % | 4.0 % |
| AssemblyAI Table 5 disease entity error rate, dictation (lower is better) | AssemblyAI Benchmarks | 11.2 % | 13.6 % |
| AssemblyAI Table 5 disease entity error rate, mean (lower is better) | AssemblyAI Benchmarks | 8.8 % | 9.2 % |
| AssemblyAI Table 5 disease entity error rate, primary care (lower is better) | AssemblyAI Benchmarks | 13.0 % | 11.8 % |
| AssemblyAI Table 5 disease entity error rate, respiratory (lower is better) | AssemblyAI Benchmarks | 2.2 % | 2.2 % |
| AssemblyAI Table 5 medication entity error rate, dictation (lower is better) | AssemblyAI Benchmarks | 20.5 % | 20.8 % |
| AssemblyAI Table 5 medication entity error rate, mean (lower is better) | AssemblyAI Benchmarks | 17.8 % | 19.8 % |
| AssemblyAI Table 5 medication entity error rate, primary care (lower is better) | AssemblyAI Benchmarks | 25.1 % | 29.3 % |
| AssemblyAI Table 5 medication entity error rate, respiratory (lower is better) | AssemblyAI Benchmarks | 7.7 % | 9.2 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Addresses (lower is better) | AssemblyAI Benchmarks | 24.8 % | 32.7 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), All classes (lower is better) | AssemblyAI Benchmarks | 25.9 % | 31.4 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), All, formatted (lower is better) | AssemblyAI Benchmarks | 42.4 % | 62.5 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Alphanumeric (lower is better) | AssemblyAI Benchmarks | 8.8 % | 10.0 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Dates (lower is better) | AssemblyAI Benchmarks | 29.3 % | 23.5 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Email (lower is better) | AssemblyAI Benchmarks | 16.3 % | 15.8 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Orgs (lower is better) | AssemblyAI Benchmarks | 25.6 % | 30.2 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Persons (lower is better) | AssemblyAI Benchmarks | 36.1 % | 49.1 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Phone (lower is better) | AssemblyAI Benchmarks | 16.9 % | 3.4 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Places (lower is better) | AssemblyAI Benchmarks | 22.3 % | 25.0 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), Products (lower is better) | AssemblyAI Benchmarks | 28.0 % | 33.0 % |
| AssemblyAI Table 6 judge-scored general English entity error (%), URLs (lower is better) | AssemblyAI Benchmarks | 20.8 % | 31.4 % |
| AssemblyAI Table 7 short-form English WER, Common Voice (lower is better) | AssemblyAI Benchmarks | 12.4 % | 7.7 % |
| AssemblyAI Table 7 short-form English WER, LibriSpeech clean (lower is better) | AssemblyAI Benchmarks | 3.3 % | 3.6 % |
| AssemblyAI Table 7 short-form English WER, LibriSpeech other (lower is better) | AssemblyAI Benchmarks | 6.7 % | 8.4 % |
| AssemblyAI Table 7 short-form English WER, mean 3 sets (lower is better) | AssemblyAI Benchmarks | 7.5 % | 6.5 % |
| AssemblyAI Table 8 accented English WER, British (lower is better) | AssemblyAI Benchmarks | 9.1 % | 9.6 % |
| AssemblyAI Table 8 accented English WER, French accent (lower is better) | AssemblyAI Benchmarks | 13.9 % | 14.8 % |
| AssemblyAI Table 8 accented English WER, Indian accent (lower is better) | AssemblyAI Benchmarks | 9.6 % | 9.6 % |
| AssemblyAI Table 8 accented English WER, mean 4 sets (lower is better) | AssemblyAI Benchmarks | 12.0 % | 12.6 % |
| AssemblyAI Table 8 accented English WER, Spanish accent (lower is better) | AssemblyAI Benchmarks | 15.2 % | 16.1 % |
| AssemblyAI Table 9 long-form English WER, CALLHOME per channel (lower is better) | AssemblyAI Benchmarks | 11.2 % | 12.1 % |
| AssemblyAI Table 9 long-form English WER, mean 3 sets (lower is better) | AssemblyAI Benchmarks | 9.8 % | 9.6 % |
| AssemblyAI Table 9 long-form English WER, Rev16 (lower is better) | AssemblyAI Benchmarks | 9.4 % | 8.1 % |
| AssemblyAI Table 9 long-form English WER, TED-LIUM (lower is better) | AssemblyAI Benchmarks | 7.4 % | 7.0 % |
| English short-form WER (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 7.5 % | 6.5 % |
| Entity error rate, report headline (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 26.1 % | 30.1 % |
| Entity error rate: codes and IDs (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 27.6 % | 46.1 % |
| Entity error rate: dates (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 43.7 % | 22.8 % |
| Entity error rate: email addresses (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 34.4 % | 40.7 % |
| Entity error rate: names (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 24.3 % | 29.0 % |
| Entity error rate: phone numbers (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 4.5 % | 11.5 % |
| Entity error rate: technical terms (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 39.4 % | 43.9 % |
| WER, report headline (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 8.6 % | 13.5 % |
| WER: heavy noise (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 9.9 % | 14.9 % |
| WER: low noise (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 8.5 % | 13.4 % |
| WER: no noise (lower is better) | AssemblyAI Universal-3.6 Pro Realtime report | 7.5 % | 12.3 % |
All data side by side
Prices are verbatim from the provider, because audio models bill per minute, character or token.