Left gets fewer words wrong. Lower has the final text ready sooner.
Soniox STT-RT v5 is ahead of GPT-4o Transcribe on both datasets together, the Pipecat Dataset, the Ocular Dataset and final-text delay.
Last updated Methodology by Dileep Chagam
Lower is better on every measure. Too close to call means the likely ranges overlap. Our guide covers testing OpenAI Realtime API voice agents.
On every clip both models were scored on, the lower WER wins, and WERs within 5 points of each other are a draw; in the rating, a wider win counts for more. Expected win rate is what the two ELO ratings predict for a match.
Short replies are decided on the error count and entities on exact capture, with no draw margin. How the rating is worked out
Soniox STT-RT v5 is more accurate on both datasets. On the 1,000 public Pipecat clips, GPT-4o Transcribe scores 6.85% and Soniox STT-RT v5 3.02%, so Soniox STT-RT v5 makes fewer errors by 3.83 points. On the 300 licensed Ocular turns, GPT-4o Transcribe scores 3.90% and Soniox STT-RT v5 2.54%, so Soniox STT-RT v5 makes fewer errors by 1.36 points. Word error rate counts incorrect, missing, and extra words, so lower is better.
Soniox STT-RT v5 is more accurate on both datasets. On the 1,000 public Pipecat clips, GPT-4o Transcribe scores 6.85% and Soniox STT-RT v5 3.02%, so Soniox STT-RT v5 makes fewer errors by 3.83 points. On the 300 licensed Ocular turns, GPT-4o Transcribe scores 3.90% and Soniox STT-RT v5 2.54%, so Soniox STT-RT v5 makes fewer errors by 1.36 points. Word error rate counts incorrect, missing, and extra words, so lower is better.
Final-text delay was not measured for one or both, so speed cannot be compared. Both are measured on the same 206 streamed turns.
Soniox STT-RT v5 makes fewer errors on real conversations, and the two are not separated on speed. A voice agent cannot reply until the final transcript arrives, and the Ocular recordings are real conversations.
Only Soniox STT-RT v5. GPT-4o Transcribe returns no text until the speaker stops, which matters if your agent reacts to partial transcripts.
Cekura runs these simulated calls on your agent and scores every one.