Left gets fewer words wrong. Lower has the final text ready sooner.
Soniox STT-RT v5 is ahead on time to first text and final-text delay, and GPT Realtime Whisper on both datasets together. The Pipecat Dataset and The Ocular Dataset are too close to call.
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
GPT Realtime Whisper is more accurate on the public clips. On the 1,000 public Pipecat clips, GPT Realtime Whisper scores 2.63% and Soniox STT-RT v5 3.02%, so GPT Realtime Whisper makes fewer errors by 0.39 points. On the 300 licensed Ocular turns, GPT Realtime Whisper scores 2.16% and Soniox STT-RT v5 2.54%, with likely ranges of 1.92 to 2.41% and 2.27 to 2.83%. The ranges overlap, so the difference is too close to call. Word error rate counts incorrect, missing, and extra words, so lower is better.
GPT Realtime Whisper is more accurate on the public clips. On the 1,000 public Pipecat clips, GPT Realtime Whisper scores 2.63% and Soniox STT-RT v5 3.02%, so GPT Realtime Whisper makes fewer errors by 0.39 points. On the 300 licensed Ocular turns, GPT Realtime Whisper scores 2.16% and Soniox STT-RT v5 2.54%, with likely ranges of 1.92 to 2.41% and 2.27 to 2.83%. The ranges overlap, so the difference is too close to call. Word error rate counts incorrect, missing, and extra words, so lower is better.
No. Soniox STT-RT v5 finishes the transcript 184ms sooner. GPT Realtime Whisper has the final transcript 1.38s after the speaker stops and Soniox STT-RT v5 1.20s, at the median. Both are measured on the same 206 streamed turns.
Soniox STT-RT v5 finishes the transcript 184ms 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.
Yes, both return partial text before the speaker stops. GPT Realtime Whisper shows first text after 3.53% and Soniox STT-RT v5 after 3.92%, at the median.
Cekura runs these simulated calls on your agent and scores every one.