AssemblyAI Universal 3.5 Pro vs Deepgram Nova-3
AssemblyAI Universal 3.5 Pro is ahead on the Pipecat Dataset and time to first text, and Deepgram Nova-3 on final-text delay. The Ocular Dataset is too close to call.
AssemblyAI Universal 3.5 Pro has a newer release. See AssemblyAI Universal 3.6 Pro vs Deepgram Nova-3.
Last updated Methodology by Dileep Chagam
Accuracy vs speed
Each dot is one model. Bottom-left is best: fewer errors and a faster final transcript.
Which to pick
Pick AssemblyAI Universal 3.5 Pro for accuracy on clean, scripted audio (1.53 points lower) and showing words while the caller is still talking (575ms sooner). Pick Deepgram Nova-3 for quick turn-taking once the caller stops (77ms sooner). For accuracy on real phone conversations, the gap is inside the margin of error, so neither is ahead.
The four metrics, side by side
- AssemblyAI Universal 3.5 Pro
- 1.93%2nd
- Deepgram Nova-3
- 3.46%12th
AssemblyAI Universal 3.5 Pro, 1.53 points lower
- AssemblyAI Universal 3.5 Pro
- 4.23%10th
- Deepgram Nova-3
- 4.39%11th
Too close to call: inside the margin of error
- AssemblyAI Universal 3.5 Pro
- 489ms2nd
- Deepgram Nova-3
- 1.06s7th
AssemblyAI Universal 3.5 Pro, 575ms sooner
- AssemblyAI Universal 3.5 Pro
- 180ms9th
- Deepgram Nova-3
- 103ms4th
Deepgram Nova-3, 77ms sooner
Lower is better on every metric. The small figure is each model's place in the whole field. What each metric measures.
- WER · Word error rate
- Share of words the model got wrong: replaced, missing or added. Lower is better.
- Pipecat Dataset
- 1,000 short public clips from Pipecat's STT benchmark dataset.
- Ocular Dataset
- 8 recordings from 4 real phone conversations, licensed from Ocular.
- TTFT · Time to first text
- How long after the audio starts streaming until the first words appear.
- TTFS · Final-text delay
- How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.
First text and final text, at three percentiles
AssemblyAI Universal 3.5 Pro shows first text after 1.08s and final text after 692ms, and Deepgram Nova-3 shows first text after 2.07s and final text after 155ms, on a bad turn (P95) across the same streamed turns.
| Measure | Model | P50 | P90 | P95 |
|---|---|---|---|---|
| Time to first text | AssemblyAI Universal 3.5 Pro | 489ms | 966ms | 1.08s |
| Deepgram Nova-3 | 1.06s | 1.10s | 2.07s | |
| Final-text delay | AssemblyAI Universal 3.5 Pro | 180ms | 532ms | 692ms |
| Deepgram Nova-3 | 103ms | 143ms | 155ms |
- P50 / P90 / P95 · Typical, slow and slowest turns
- Half of turns are faster than P50. Only 1 in 10 is slower than P90, and 1 in 20 slower than P95.
How each ranks among all 17 models
Place among the 17 models on each metric. Lower is better on all four.
| Metric | Universal 3.5 Pro | Nova-3 |
|---|---|---|
| Pipecat Dataset WER | 1.93%2nd of 17 | 3.46%12th of 17 |
| Ocular Dataset WER | 4.23%10th of 17 | 4.39%11th of 17 |
| Time to first text | 489ms2nd of 15 | 1.06s7th of 15 |
| Final-text delay | 180ms9th of 17 | 103ms4th of 17 |
How much audio each model was tested on
| Model | Pipecat clips | Ocular recordings | Turns timed | Failed |
|---|---|---|---|---|
| AssemblyAI Universal 3.5 Pro | 1,000 | 8 | 205 | 1 |
| Deepgram Nova-3 | 1,000 | 8 | 202 | 0 |
Only clips with usable text count, so the two totals can differ. How WER is scored.
Related comparisons
Both providers, every model: AssemblyAI vs Deepgram
- AssemblyAI Universal 3.5 Pro vs AssemblyAI Universal 3.6 Pro
- Deepgram Nova-3 vs GPT-4o Mini Transcribe
- Deepgram Nova-3 vs Smallest Pulse
- Deepgram Nova-3 vs GPT-4o Transcribe
- Deepgram Nova-3 vs Deepgram Flux English
- Deepgram Nova-3 vs Inworld STT-1
- Deepgram Nova-3 vs Soniox STT-RT v5
- Deepgram Nova-3 vs Speechmatics Linden
- Deepgram Nova-3 vs Cartesia Ink 2
- Deepgram Nova-3 vs GPT Realtime Whisper
- Deepgram Nova-3 vs Reson8
- Deepgram Nova-3 vs Google Chirp 3
- Deepgram Nova-3 vs Deepgram Flux Multilingual
- Deepgram Nova-3 vs AssemblyAI Universal 3.6 Pro
- Deepgram Nova-3 vs Gradium
Frequently asked questions
Which is more accurate, AssemblyAI Universal 3.5 Pro or Deepgram Nova-3?
AssemblyAI Universal 3.5 Pro is more accurate on the public clips. On the 1,000 public Pipecat clips, AssemblyAI Universal 3.5 Pro scores 1.93% and Deepgram Nova-3 3.46%, so AssemblyAI Universal 3.5 Pro makes fewer errors by 1.53 points. On the eight licensed Ocular recordings, AssemblyAI Universal 3.5 Pro scores 4.23% and Deepgram Nova-3 4.39%, with likely ranges of 3.57 to 4.68% and 3.42 to 5.25%. The ranges overlap, so the difference is too close to call. Word error rate counts incorrect, missing, and extra words, so lower is better.
Is AssemblyAI Universal 3.5 Pro faster than Deepgram Nova-3?
No. Deepgram Nova-3 finishes the transcript 77ms sooner. AssemblyAI Universal 3.5 Pro has the final transcript 180ms after the speaker stops and Deepgram Nova-3 103ms, at the median. Both are measured on the same 206 streamed turns.
Is AssemblyAI Universal 3.5 Pro or Deepgram Nova-3 better for voice agents?
Deepgram Nova-3 finishes the transcript 77ms sooner, and the two are too close to call on real conversations. A voice agent cannot reply until the final transcript arrives, and the Ocular recordings are real conversations.
Do AssemblyAI Universal 3.5 Pro and Deepgram Nova-3 stream text while the speaker is talking?
Yes, both return partial text before the speaker stops. AssemblyAI Universal 3.5 Pro shows first text after 489ms and Deepgram Nova-3 after 1.06s, at the median, so AssemblyAI Universal 3.5 Pro is 575ms sooner.