whatmodel.sh / stt / Nova-3
Nova-3
↗ share
+ contribute
Nova-3 is a closed-weights speech-to-text model from Deepgram, released February 2025. Licence: proprietary.
Score within the Open ASR English short-form WER (March 2026) comparison set only.
$ how it compares
The speech-to-text models ranked either side of Nova-3, out of 176 with a score, then more from Deepgram.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 25 |  Cohere Transcribe (03-2026) | 62±2 | api $0.0042/audio min +local | ||
| 26 |  Granite Speech 4.1 2B NAR | 62±4 | local | ||
| 27 | Nova-3 | 61±2 | api $0.0052/audio min | ||
| 28 |  Canary-Qwen-2.5B | 61±2 | local | ||
| 29 |  Pulse (STT) | 61±3 | api $0.0040/audio min | ||
| more from Deepgram | |||||
| 43 | Deepgram Nova 2 | 60±4 | api $0.0058/audio min | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
multilingual transcription▂··▃▄█▆▅▄▂·▂ #10 of 36 * · 12% of its score · 2 tests
Beats Cohere Transcribe Arabic (07-2026) here, though they rank higher overall. Trails SeamlessM4T v2 Large, Whisper large-v3, Omnilingual ASR LLM 7B v2 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CoSE-E enterprise code-switching · 4 European pairs mean WER | 5.9% ✎ | █··██···█··█ beats 2/4 | 8.1% |
| mu-bench voice-agent caller utterances · batch UER (5 locales) | 28.0% ✎ | █·····▄····▄ beats 0/3 | 3.6% |
other tests17 not in a task · 87% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Pipecat STT Benchmark · pooled semantic WER (1000 smart-turn samples, Sonnet 5.5 judge) | 1.4% ✎ | ▃·▃▅·▅▅·▃▅██ beats 8/16 | 22% |
| Bangla STT benchmark · BanSpeech 1,001-clip batch CER | 10.1% ✎ | █·▄··▄··▄·▄▄ beats 4/6 | 11% |
| RW-Voice-EQ ASR · 4 real-world English sets, mean WER | 9.5% ✎ | █▃▃▅▃···▃··▃ beats 1/9 | 6.8% |
| VoiceCodeBench · batch canonical token/entity match (CTEM) | 90.9% ✎ | ▄···▄··▄▄▄·█ beats 5/6 | 6.5% |
| verli STT benchmark · LibriSpeech test-clean WER | 2.5% ✎ | ▃·····▃▅█▃▅▅ beats 1/11 | 5.0% |
| GigaSpeechBench · 7 Arabic dialect WER mean | 46.6% ✎ | █··█·██··█·█ beats 3/5 | 5.0% |
| Russian STT · podlodka corpus WER (20-clip pilot) | 7.1% ✎ | █··········▄ beats 2/2 | 4.9% |
| Russian STT · rulibrispeech corpus WER | 8.9% ✎ | █··█·······█ beats 2/2 | 4.9% |
9 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VoiceCodeBench · batch WER (300 English workplace segments) | 8.2% ✎ | ▄·▄·▄·▄█···▄ beats 1/6 | 4.7% |
| Argmax OpenBench · earnings22-keywords WER (no keyword prompting) | 15.3% ✎ | █··········▄ beats 1/2 | 4.3% |
| Russian STT · fleurs corpus WER | 6.6% ✎ | █··█·······█ beats 0/2 | 3.7% |
| koedesk FLEURS first-200 · 9-language headline error mean | 7.5% ✎ | ▄·····▄··▄█▄ beats 0/5 | 2.9% |
| verli STT benchmark · SPGISpeech 1,200-clip sample WER | 2.9% ✎ | ▂·····▄▂█▂▂▄ beats 7/11 | 2.4% |
| verli STT benchmark · Earnings-22 WER (2,741 utt.) | 16.1% ✎ | ▃·▃▅··▃▃▅▅█▃ beats 1/13 | 2.2% |
| Quranic ASR Leaderboard · held-out recitation (600 clips) overall WER | 14.1% ✎ | ▃·▅·▅▃▃·▃▃▃█ beats 8/12 | 1.5% |
| Voice memos · RU 10-clip pilot median WER | 31.0% ✎ | █··········█ beats 0/1 | — |
| Voice memos · EN 10-clip pilot median WER | 23.6% ✎ | █··········█ beats 0/1 | — |
$ where to run it
Through an API
api · Cloudflare · $0.0052/audio min · developers.cloudflare.com ↗ · price sourceCloudflare Workers AI published USD per input audio minute; file transcription, not WebSocket streaming. Plan base fees excluded. USD table is rounded; backend bills Neurons. Product: Workers AI file transcription. API model: @cf/deepgram/nova-3. Settings: streaming=no, plan=paid, free allocation=excluded. Price basis: published rounded USD/audio-minute. Currency: USD.
$ how the score was worked out · Pipecat STT Benchmark · seman… 22%, Russian STT benchmark by doma… 14%, VoiceCodeBench (arXiv:2608.28… 11% and 9 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
not counted (2)
| Live voice-memo STT pilot | Voice memos · RU 10-clip pilot median WER [correct] | 31.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Live voice-memo STT pilot | Voice memos · EN 10-clip pilot median WER [correct] | 23.6% | fewer than three models have this test yet, or this source ran it for this model alone |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json