Skip to main content

DeepgramAudio modelActive

Nova-3

PopularityRank 101 of 631

51out of 100same as last month

based on two data groups, mostly search

Summary

Nova-3 is Deepgram’s general-purpose speech-to-text model for batch and streaming. Its multilingual option covers English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch, with English as the default. Deepgram documents meetings, captions, multi-speaker recordings, and noisy or far-field audio for the profile.

Nova-3 transcribes speech but does not provide model-integrated turn detection. Flux adds that end-of-turn detection for English voice agents.

GraduallyBench Audio

Speech recognition compared automatically.

The pilot uses the same 32 read-speech recordings in German and English, plus four digital-silence clips. The comparison uses source transcripts without human verification. Each language contains four speakers. Spontaneous speech and background noise are outside this cohort.

Word error rate (WER) counts replaced, omitted and inserted words relative to the reference words. Character error rate (CER) compares characters. Lower values mean fewer differences from the reference. WER can exceed 100 %. Counts below each rate show errors / reference units. Silence is evaluated separately.

Protocol audio-stt-read-speech-automated-v1. Actual provider billing is unknown.

Deepgram Nova 3

Cloudflare Workers AI · @cf/deepgram/nova-3 · Tested Nov 5, 2026

Successful requests 36 / 36. The exact backend version is not confirmed by the requested route.

Read-speech transcription
LanguageSuccessful responsesWERCERWER, speaker mean
German16 / 168.55 %47 / 5502.61 %85 / 32628.16 %n = 4 / 4
English16 / 162.04 %10 / 4900.66 %17 / 25802.02 %n = 4 / 4
Digital silence
Language settingClips with invented speechInvented words
DE0 / 20
EN0 / 20

Request duration, median / p95 1,180.5 / 2,849 ms (n = 36). Includes network time and client processing.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 307 published benchmark results

Page 1 of 26

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Artificial Analysis AA-WER (non-streaming) (Show measurement, test conditions, and source)
Source value
5.2
Score
5.2
Metric
word error rate (WER)
Unit
%
Benchmark version
AA-WER v2.2
Category
audio
Direction
Lower is better
Source type
independent-evaluator
Evaluator
Artificial Analysis
Status
active
Retrieved at
2026-10-15
Methodology
Independent evaluation. Artificial Analysis directly tested the named API endpoint; audio-duration-weighted WER across AA-AgentTalk (50%), VoxPopuli-Cleaned-AA English (25%), and Earnings22-Cleaned-AA (25%), about eight hours total. Current normalizer AA-WER v2.2.
Source Artificial Analysis
Oruk IEMOCAP transcript-sentiment valence accuracy (Show measurement, test conditions, and source)
Source value
46.3
Score
46.3
Metric
three-class sentiment-polarity accuracy
Unit
%
Benchmark version
Oruk report revision 2026-09-06; evaluated 2026-08-01
Category
audio
Direction
Higher is better
Sample
1947
Lower confidence bound
44.1
Upper confidence bound
48.5
Source type
vendor-reported
Evaluator
Oruk diagnostic report
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Deepgram nova-3 transcript sentiment; IEMOCAP acted speech, evaluated 2026-08-01; three-way negative/neutral/positive mapping of gold labels and native sentiment; 1947 scored responses from the 2008-clip test split.
Context
The publisher sells a competing product and has a direct commercial conflict; this is not independent third-party validation. The input is transcript sentiment, not acoustic emotion recognition, and the metric is not word error rate. The sarcasm rows measure an extra trained probe and its data split, not a native Nova-3 sarcasm classifier or general transcription quality.
Source Oruk diagnostic report
Oruk MUStARD speaker-dependent sarcasm macro F1 (Show measurement, test conditions, and source)
Source value
51.7
Score
51.7
Metric
macro F1 multiplied by 100
Unit
%
Benchmark version
Oruk report revision 2026-09-06; evaluated 2026-08-01
Category
audio
Direction
Higher is better
Sample
690
Source type
vendor-reported
Evaluator
Oruk diagnostic report
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Deepgram nova-3 transcript sentiment with an additional trained logistic probe for sarcasm; MUStARD speaker-dependent standard 5-fold cross-validation, evaluated 2026-08-01; out-of-fold predictions over 690 utterances.
Context
The publisher sells a competing product and has a direct commercial conflict; this is not independent third-party validation. The input is transcript sentiment, not acoustic emotion recognition, and the metric is not word error rate. The sarcasm rows measure an extra trained probe and its data split, not a native Nova-3 sarcasm classifier or general transcription quality.
Source Oruk diagnostic report
Oruk MUStARD speaker-independent sarcasm macro F1 (Show measurement, test conditions, and source)
Source value
39.5
Score
39.5
Metric
macro F1 multiplied by 100
Unit
%
Benchmark version
Oruk report revision 2026-09-06; evaluated 2026-08-01
Category
audio
Direction
Higher is better
Sample
690
Source type
vendor-reported
Evaluator
Oruk diagnostic report
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Deepgram nova-3 transcript sentiment with an additional trained logistic probe for sarcasm; MUStARD speaker-independent grouped 5-fold cross-validation, evaluated 2026-08-01; speakers do not overlap between training and test folds, 690 utterances.
Context
The publisher sells a competing product and has a direct commercial conflict; this is not independent third-party validation. The input is transcript sentiment, not acoustic emotion recognition, and the metric is not word error rate. The sarcasm rows measure an extra trained probe and its data split, not a native Nova-3 sarcasm classifier or general transcription quality.
Source Oruk diagnostic report
Oruk MUStARD cross-show sarcasm macro F1 (Show measurement, test conditions, and source)
Source value
29.9
Score
29.9
Metric
macro F1 multiplied by 100
Unit
%
Benchmark version
Oruk report revision 2026-09-06; evaluated 2026-08-01
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
Oruk diagnostic report
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Deepgram nova-3 transcript sentiment with an additional trained logistic probe for sarcasm; MUStARD cross-show evaluation on Friends after training on The Big Bang Theory, Golden Girls and Sarcasmoholics, evaluated 2026-08-01. The scored test population is not supplied.
Context
The publisher sells a competing product and has a direct commercial conflict; this is not independent third-party validation. The input is transcript sentiment, not acoustic emotion recognition, and the metric is not word error rate. The sarcasm rows measure an extra trained probe and its data split, not a native Nova-3 sarcasm classifier or general transcription quality. The cross-show sample count is unknown; the full-dataset count is not assigned to this split.
Source Oruk diagnostic report
Coval rolling-30-day pooled word error rate (Nova 3 endpoint) (Show measurement, test conditions, and source)
Source value
5.304104762675151
Score
5.304104762675151
Metric
Pooled word error rate across Coval's fixed public speech-recognition audio conditions
Unit
%
Benchmark version
Coval rolling 30-day snapshot, last measured 2026-10-16, source response fetched 2026-10-16T15:36:04.059740Z
Category
audio
Direction
Lower is better
Sample
26025
Source type
independent-evaluator
Evaluator
Coval
Status
superseded
Retrieved at
2026-10-16
Methodology
Independent evaluation. Exact hosted model ID nova-3 from Deepgram. Coval reports daily fixed-public-audio endpoint tests over a rolling 30-day window, last measured 2026-10-16. Stored score is pooled_value=5.304104762675151 from the WER row, not mean_value; Coval's visible 5.3% is rounded to one decimal. The reported WER sample count comes from the same row.
Context
The published WER sample count covers only scored items in that metric row, not all attempted requests or failures. No confidence interval is published. This rolling Coval endpoint result is distinct from Artificial Analysis AA-WER.
Source Coval
Cekura Converse-STT, Deepgram Nova-3 Pipecat WER (Show measurement, test conditions, and source)
Source value
3.46
Score
3.46
Metric
word error rate (WER) on Pipecat public short-clip speech-to-text set
Unit
%
Benchmark version
Cekura Converse-STT page last updated 2026-09-16, retrieved 2026-10-17
Category
audio
Direction
Lower is better
Sample
1000
Lower confidence bound
3.16
Upper confidence bound
3.78
Source type
independent-evaluator
Evaluator
Cekura Bench
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Streaming STT model row named Deepgram Nova-3. The Pipecat dataset contains 1,000 public clips; the row reports 1,000 usable model outputs. Cekura labels the uncertainty estimate a likely 95% range. The served endpoint build is not pinned.
Context
The published interval is Cekura's likely 95% range, not a separately documented model confidence interval. The page states that per-model counts are usable outputs. This distinct clean public-clip WER observation is not the existing AA-WER or Coval WER row.
Source Cekura Bench
Cekura Converse-STT, Deepgram Nova-3 Ocular WER (Show measurement, test conditions, and source)
Source value
4.39
Score
4.39
Metric
word error rate (WER) on Ocular licensed real-phone-conversation speech-to-text set
Unit
%
Benchmark version
Cekura Converse-STT page last updated 2026-09-16, retrieved 2026-10-17
Category
audio
Direction
Lower is better
Sample
8
Lower confidence bound
3.42
Upper confidence bound
5.25
Source type
independent-evaluator
Evaluator
Cekura Bench
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Streaming STT model row named Deepgram Nova-3. Ocular provides eight licensed recordings from four real phone conversations; the row reports usable outputs for all eight. Cekura labels the uncertainty estimate a likely 95% range. The served endpoint build is not pinned.
Context
The published interval is Cekura's likely 95% range, not a separately documented model confidence interval. The page states that per-model counts are usable outputs. This distinct real-call WER observation is not the existing AA-WER or Coval WER row.
Source Cekura Bench
Commercial ASR on Dysarthric Speech (TORGO) (Show measurement, test conditions, and source)
Source value
2.77
Score
2.77
Metric
Mild dysarthria word error rate (WER), utterance-weighted mean
Unit
%
Benchmark version
arXiv:2512.17474v2; API evaluation 2025-09-17 to 2025-09-18
Category
dysarthric-speech-recognition
Direction
Lower is better
Source type
independent-evaluator
Evaluator
Benchmarking Commercial Speech Recognition on Dysarthric Speech
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact paper row Deepgram Nova-3, model ID nova-3; accessed via public API during 2025-09-17 to 2025-09-18. Conventional ASR used default API parameters with no fine-tuning or speaker adaptation; this row is the default non-prompted condition. TORGO head-mounted-microphone audio was supplied as 16 kHz mono WAV; utterances shorter than three words and records without reference transcripts were excluded.
Context
Descriptive severity-stratified result from eight TORGO speakers, not general English ASR performance. The paper does not publish model-specific accepted-output N, confidence intervals, or an immutable provider serving revision. The reported ± value is standard deviation, not a CI.
Source Benchmarking Commercial Speech Recognition on Dysarthric Speech
Commercial ASR on Dysarthric Speech (TORGO) (Show measurement, test conditions, and source)
Source value
1.44
Score
1.44
Metric
Mild dysarthria character error rate (CER), utterance-weighted mean
Unit
%
Benchmark version
arXiv:2512.17474v2; API evaluation 2025-09-17 to 2025-09-18
Category
dysarthric-speech-recognition
Direction
Lower is better
Source type
independent-evaluator
Evaluator
Benchmarking Commercial Speech Recognition on Dysarthric Speech
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact paper row Deepgram Nova-3, model ID nova-3; accessed via public API during 2025-09-17 to 2025-09-18. Conventional ASR used default API parameters with no fine-tuning or speaker adaptation; this row is the default non-prompted condition. TORGO head-mounted-microphone audio was supplied as 16 kHz mono WAV; utterances shorter than three words and records without reference transcripts were excluded.
Context
Descriptive severity-stratified result from eight TORGO speakers, not general English ASR performance. The paper does not publish model-specific accepted-output N, confidence intervals, or an immutable provider serving revision. The reported ± value is standard deviation, not a CI.
Source Benchmarking Commercial Speech Recognition on Dysarthric Speech
Commercial ASR on Dysarthric Speech (TORGO) (Show measurement, test conditions, and source)
Source value
98.54
Score
98.54
Metric
Mild dysarthria semantic preservation score (SeMaScore), utterance-weighted mean
Unit
%
Benchmark version
arXiv:2512.17474v2; API evaluation 2025-09-17 to 2025-09-18
Category
dysarthric-speech-recognition
Direction
Higher is better
Source type
independent-evaluator
Evaluator
Benchmarking Commercial Speech Recognition on Dysarthric Speech
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact paper row Deepgram Nova-3, model ID nova-3; accessed via public API during 2025-09-17 to 2025-09-18. Conventional ASR used default API parameters with no fine-tuning or speaker adaptation; this row is the default non-prompted condition. TORGO head-mounted-microphone audio was supplied as 16 kHz mono WAV; utterances shorter than three words and records without reference transcripts were excluded.
Context
Descriptive severity-stratified result from eight TORGO speakers, not general English ASR performance. The paper does not publish model-specific accepted-output N, confidence intervals, or an immutable provider serving revision. The reported ± value is standard deviation, not a CI.
Source Benchmarking Commercial Speech Recognition on Dysarthric Speech
Commercial ASR on Dysarthric Speech (TORGO) (Show measurement, test conditions, and source)
Source value
97.53
Score
97.53
Metric
Mild dysarthria BERTScore, utterance-weighted mean
Unit
%
Benchmark version
arXiv:2512.17474v2; API evaluation 2025-09-17 to 2025-09-18
Category
dysarthric-speech-recognition
Direction
Higher is better
Source type
independent-evaluator
Evaluator
Benchmarking Commercial Speech Recognition on Dysarthric Speech
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact paper row Deepgram Nova-3, model ID nova-3; accessed via public API during 2025-09-17 to 2025-09-18. Conventional ASR used default API parameters with no fine-tuning or speaker adaptation; this row is the default non-prompted condition. TORGO head-mounted-microphone audio was supplied as 16 kHz mono WAV; utterances shorter than three words and records without reference transcripts were excluded.
Context
Descriptive severity-stratified result from eight TORGO speakers, not general English ASR performance. The paper does not publish model-specific accepted-output N, confidence intervals, or an immutable provider serving revision. The reported ± value is standard deviation, not a CI.
Source Benchmarking Commercial Speech Recognition on Dysarthric Speech

Page 1 of 26

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 45 models

pyannote Community-173
Nemotron-3-Diarization57
Gemini 3.5 Transcribe52
Confucius4-R2T252
NVIDIA Nemotron 3.5 Transcribe52
Granite Speech 5.0 470M TurboCTC52
Kokoro-82M51
Nova-351
Fish Audio S2 Pro51
Granite Speech 5.0 470M TurboCTC NC50
Orukeet49
Lyria 3.547
Gemini 3.5 Live Translate46
Phonon-244
Aura-244
Gemini 3.1 Flash TTS44
MiniMax-Music341
Audio8-ASR-Infinite41
GPT-Realtime 2.141
Maya 2 Native40

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 7, 2026

Nova-3NVIDIA Nemotron 3.5 Transcribe 52Granite Speech 5.0 470M TurboCTC 52Kokoro-82M 51Fish Audio S2 Pro 51

Method changes in the timeline, a jump at these points is no market event:

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

Nova-3 compared

Similar models

More models of the same kind as Nova-3, chosen by model line, developer, size, price and release date.

All specifications

Architecture

Model class
Speech recognition
Source: Model class

Input and output

Input
Audio file or stream
Source: Input
Output
Text
Source: Output
Languages
Multilingual option for English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch; English is the default
Source: Languages

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
API
Source: Access

API

API model ID
nova-3
Source: API model ID

Further details

Task
Transcription
Source: Task
Supported details
Multilingual, batch and streaming
Source: Supported details
Intended use
General-purpose speech recognition without integrated turn detection
Source: Intended use

Sources and data date

Show all 46 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

EvaluatorTest seriesComposite indicesRecorded active measurements
Artificial Analysis001

Without vendor claims and historical or disputed results; those stay labelled in the single values. 1 measurements are not yet assigned to a test series.

TypeEvidence and data date
Research statusResearch date October 15, 2026. 42 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published

torgo-model-specific-effective-sample-count

Original wording

The paper gives corpus counts of 195 mild, 253 moderate, and 284 severe utterances, but not the exact successful-output N used for each Nova-3 metric. No sample count is assigned.

Not published

torgo-confidence-intervals

Original wording

Table 3 reports mean ± standard deviation only. It does not give confidence-interval bounds.

Not published

torgo-nova-3-serving-revision

Original wording

The study identifies API model ID nova-3 and an evaluation window in September 2025, but no immutable provider build or checkpoint hash.

Unresolved

exact-serving-snapshot-and-unpublished-method-details

Note

The captured external source reports a benchmark-specific model or hosted route but does not establish an immutable currently served checkpoint. Unpublished row-level denominators and confidence intervals remain unknown; sample counts were entered only for cells that state them.

Unresolved

confidence intervals and per-model N

Note

AirCaps reports point estimates only and no CIs. VoiceCodeBench specifies 300 dataset clips but does not separately report model-specific evaluated N; leaderboard latency method and direction are unspecified.

Additional sourceDeepgram speech-to-text documentation (retrieved September 24, 2026; October 15, 2026; October 16, 2026; October 17, 2026) · Deepgram speech-to-text models and languages · Editorial description reviewed October 2, 2026
Additional sourceArtificial Analysis (retrieved October 15, 2026)
Additional sourceOruk diagnostic report (retrieved October 16, 2026)
Additional sourceCoval (retrieved October 16, 2026; October 17, 2026; October 18, 2026; October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 5, 2026; November 6, 2026)
Additional sourceCekura Bench (retrieved October 17, 2026; October 23, 2026)
Additional sourceBenchmarking Commercial Speech Recognition on Dysarthric Speech (retrieved October 17, 2026)
Additional sourceSpeko independent speech benchmark (retrieved October 17, 2026)
Additional sourceAssemblyAI Universal-3.6 Pro Realtime report (retrieved October 17, 2026)
Additional sourceSuperwhisper speech benchmark (retrieved October 17, 2026)
Additional sourceMU-Bench (retrieved October 17, 2026)
Additional sourceMenta, The TTS-STT Flywheel: Synthetic Entity-Dense Audio Closes the Indic ASR Gap Where Commercial and Open-Source Systems Fail (arXiv:2605.03073v1) (retrieved October 17, 2026)
Additional sourceDeepgram model comparison and internal test suite (retrieved October 17, 2026)
Additional sourceOcular Converse-STT Benchmark (retrieved October 18, 2026)
Additional sourceVoice Arena Diarization-Bench v1 (retrieved October 18, 2026)
Additional sourceNyra Verbatim Speech Benchmark (retrieved October 18, 2026)
Additional sourceSpecific Labs (retrieved October 19, 2026)
Additional sourceAirCaps A5Sv2 ASR benchmark (retrieved October 19, 2026)
Additional sourceVals AI / BeSimple AI VoiceCodeBench (retrieved October 19, 2026)
Additional sourceAssemblyAI Benchmarks (retrieved October 19, 2026)
Additional sourceDeepgram German Benchmarks (retrieved October 17, 2026; October 22, 2026)
Additional sourcePerit PERIT-STT speech-to-text leaderboard (retrieved October 23, 2026)
Additional sourceModel metadata (retrieved October 15, 2026)
Additional sourceModel metadata (retrieved October 16, 2026)
Additional sourceModel metadata (retrieved October 17, 2026)1234
Additional sourceModel metadata (retrieved October 17, 2026; October 28, 2026)
Additional sourceModel metadata (retrieved October 18, 2026)12
Additional sourceModel metadata (retrieved October 28, 2026)1234
Additional sourceModel metadata (retrieved November 5, 2026)123
Additional sourceModel metadata (retrieved November 6, 2026)123
Additional sourceOpenRouter: Popularity index (derived), CC BY 4.0
Additional sourceVercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional sourceOllama: cumulative downloads for the popularity index
Additional sourceReplicate: cumulative runs for the popularity index
Additional sourceGoogle Trending Now: limited feed selection for Germany and the US
Additional sourceCreative Commons Attribution 4.0