Deepgram Nova-3 vs Soniox STT-RT v5
Soniox STT-RT v5 is ahead on the Pipecat Dataset, the Ocular Dataset and final-text delay, and Deepgram Nova-3 on time to first text.
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%12th
- Soniox STT-RT v5
- 2.54%8th
Soniox STT-RT v5, 0.92 points lower
Word error rate on eight licensed recordings provided by Ocular.
- Deepgram Nova-3
- 4.39%11th
- Soniox STT-RT v5
- 3.92%6th
Soniox STT-RT v5, 0.47 points lower
Median delay before the first words arrive, from the first audio packet.
- Deepgram Nova-3
- 1.06s7th
- Soniox STT-RT v5
- 1.20s9th
Deepgram Nova-3, 131ms sooner
Median delay after the speaker stops before the transcript is final.
- Deepgram Nova-3
- 103ms4th
- Soniox STT-RT v5
- 38ms1st
Soniox STT-RT v5, 65ms 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 |
| Soniox STT-RT v5 | 1.20s | 1.67s | 1.71s | |
| Final-text delay | Deepgram Nova-3 | 103ms | 143ms | 155ms |
| Soniox STT-RT v5 | 38ms | 72ms | 79ms |
Both against every other model
Open on the metric where the two sit furthest apart, with the other three a tab away.
Word error rate on eight licensed recordings provided by Ocular. Lower is better.
| # | Model | Ocular Dataset | clips | |
|---|---|---|---|---|
| 1 | Reson8 | 2.93% | 8 | |
| 2 | Cartesia Ink 2 | 3.09% | 8 | |
| 3 | Smallest Pulse | 3.46% | 8 | |
| 4 | GPT Realtime Whisper | 3.53% | 8 | |
| 5 | Google Chirp 2 | 3.80% | 8 | |
| 6 | Soniox STT-RT v5 | 3.92% | 8 | |
| 7 | Speechmatics Linden | 3.93% | 8 | |
| 8 | Deepgram Flux English | 4.18% | 8 | |
| 9 | Google Chirp 3 | 4.19% | 8 | |
| 10 | AssemblyAI Universal 3.5 Pro | 4.23% | 8 | |
| 11 | Deepgram Nova-3 | 4.39% | 8 | |
| 12 | Gradium | 4.45% | 8 | |
| 13 | GPT-4o Mini Transcribe | 4.45% | 8 | |
| 14 | AssemblyAI Universal 3.6 Pro | 4.50% | 8 | |
| 15 | Deepgram Flux Multilingual | 4.53% | 8 | |
| 16 | Inworld STT-1 | 10.60% | 8 | |
| 17 | GPT-4o Transcribe | 12.45% | 8 |
Counts are the clips each model returned usable text for, so they differ between models and between datasets.
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 |
| Soniox STT-RT v5 | 1,000 | 8 | 200 | 6 |
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 Soniox STT-RT v5 more accurate?
On the 1,000 public Pipecat clips, Deepgram Nova-3 scores 3.46% and Soniox STT-RT v5 2.54%, so Soniox STT-RT v5 makes fewer errors by 0.92 points. On the eight licensed Ocular recordings, Deepgram Nova-3 scores 4.39% and Soniox STT-RT v5 3.92%, so Soniox STT-RT v5 makes fewer errors by 0.47 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 Soniox STT-RT v5?
Deepgram Nova-3 returns first text 131ms sooner, at 1.06s against 1.20s at the median. Soniox STT-RT v5 has the final transcript 65ms sooner, 38ms after the speaker stops against 103ms. Both are measured on the same 206 streamed turns.
How were Deepgram Nova-3 and Soniox STT-RT v5 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 Soniox STT-RT v5 for 1,000. Latency uses each provider's own supported finalization contract.