Deepgram Nova-3 vs GPT-4o Transcribe
Deepgram Nova-3 is ahead of GPT-4o Transcribe on the Pipecat Dataset, the Ocular Dataset and final-text delay.
Quality and speed frontier
Every model's error rate on the public clips, against how long its transcript takes to settle.
The four metrics, side by side
Word error rate on 1,000 clips from Pipecat's STT benchmark dataset.
- Deepgram Nova-3
- 3.46%10th
- GPT-4o Transcribe
- 3.90%12th
Deepgram Nova-3, 0.44 points lower
Word error rate on eight licensed recordings provided by Ocular.
- Deepgram Nova-3
- 4.39%10th
- GPT-4o Transcribe
- 12.45%15th
Deepgram Nova-3, 8.06 points lower
Median delay before the first words arrive, from the first audio packet.
- Deepgram Nova-3
- 1.06s6th
- GPT-4o Transcribe
- No text until the speaker stops
Not comparable
Median delay after the speaker stops before the transcript is final.
- Deepgram Nova-3
- 103ms2nd
- GPT-4o Transcribe
- 652ms15th
Deepgram Nova-3, 549ms sooner
Lower is better on every metric. The small figure is each model's place in the whole field.
First text and final text, at three percentiles
Both measured on the same 206 streamed Ocular Dataset turns. P90 and P95 say what a caller meets on a bad turn, which a median hides.
| Measure | Model | P50 | P90 | P95 |
|---|---|---|---|---|
| Time to first text | Deepgram Nova-3 | 1.06s | 1.10s | 2.07s |
| GPT-4o Transcribe | No text until the speaker stops | |||
| Final-text delay | Deepgram Nova-3 | 103ms | 143ms | 155ms |
| GPT-4o Transcribe | 652ms | 1.73s | 1.97s | |
Both against every other model
Open on the metric where the two sit furthest apart, with the other three a tab away.
Final-text delay after the speaker stops. Lower is better.
| # | Model | Final text | turns | |
|---|---|---|---|---|
| 1 | Inworld STT-1 | 41ms | 206 | |
| 2 | Deepgram Nova-3 | 103ms | 202 | |
| 3 | Deepgram Flux English | 108ms | 199 | |
| 4 | Deepgram Flux Multilingual | 114ms | 175 | |
| 5 | Cartesia Ink 2 | 123ms | 198 | |
| 6 | Speechmatics Linden | 147ms | 194 | |
| 7 | AssemblyAI Universal 3.5 Pro | 180ms | 205 | |
| 8 | Smallest Pulse | 201ms | 206 | |
| 9 | Gradium | 212ms | 198 | |
| 10 | Reson8 | 235ms | 201 | |
| 11 | Google Chirp 3 | 475ms | 206 | |
| 12 | GPT Realtime Whisper | 543ms | 206 | |
| 13 | Google Chirp 2 | 568ms | 205 | |
| 14 | GPT-4o Mini Transcribe | 630ms | 206 | |
| 15 | GPT-4o Transcribe | 652ms | 206 |
Median across the same 206 Ocular Dataset turns. Providers finalize differently, so compare models with similar finalization contracts before treating small gaps as meaningful.
Both on the other two views
Pipecat vs Ocular accuracy
Every model's error rate on the public clips, against the same model on real conversations.
Transcription response time
How soon each model returns its first words, against how soon its transcript is final.
What the numbers rest on
| Model | Pipecat clips | Ocular recordings | Turns timed | Failed |
|---|---|---|---|---|
| Deepgram Nova-3 | 1,000 | 8 | 202 | 0 |
| GPT-4o Transcribe | 1,000 | 8 | 206 | 0 |
A clip counts only where the model returned usable text, so counts differ between models. Word error rate is measured the same way throughout: incorrect, missing, and extra words against the reference transcript.
Frequently asked questions
Is Deepgram Nova-3 or GPT-4o Transcribe more accurate?
On the 1,000 public Pipecat clips, Deepgram Nova-3 scores 3.46% and GPT-4o Transcribe 3.90%, so Deepgram Nova-3 makes fewer errors by 0.44 points. On the eight licensed Ocular recordings, Deepgram Nova-3 scores 4.39% and GPT-4o Transcribe 12.45%, so Deepgram Nova-3 makes fewer errors by 8.06 points. Word error rate counts incorrect, missing, and extra words against the reference transcript, so lower is better.
Which is faster, Deepgram Nova-3 or GPT-4o Transcribe?
GPT-4o Transcribe returns no text while the speaker is talking, so time to first text cannot be compared. Deepgram Nova-3 has the final transcript 549ms sooner, 103ms after the speaker stops against 652ms. Both are measured on the same 206 streamed turns.
How were Deepgram Nova-3 and GPT-4o Transcribe compared?
Both ran the same two datasets under the same conditions as every other model in the benchmark: 1,000 public clips from Pipecat's STT benchmark dataset and eight licensed conversation recordings from Ocular. Deepgram Nova-3 returned usable text for 1,000 of the public clips and GPT-4o Transcribe for 1,000. Latency uses each provider's own supported finalization contract.