Lower and further left is better · Nemotron 3.5 Transcribe doesn't support streaming, so it is not shown
AssemblyAI STT 3.7 Realtime Agent is 2nd of 24 on the Pipecat Dataset and 12th of 24 on the Ocular Dataset.
Last updated Methodology by Dileep Chagam
Lower is better on every measure. Hover a column for that model's full numbers.
On real calls, dropped words are its biggest source of error. On clean clips, wrong words lead.
67%
of its errors on the Ocular calls are words it left out.
Median across all 24 models: 51%
Mix of errors
Real calls (Ocular Dataset)
Clean clips (Pipecat Dataset)
Timed on the same 206 streamed Ocular turns as every other model. P90 and P95 show what a caller meets on a slow turn.
How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.
How long after the audio starts streaming until the first words appear.
Every other model's typical turn, on the same axis.
1,000 Pipecat turns · 300 Ocular turns · 206 turns timed · 0 failed requests
AssemblyAI STT 3.7 Realtime Agent places highest on the Pipecat Dataset, 2nd of 24 at 1.68%. Its weakest placing is the Ocular Dataset, 12th of 24 at 3.81%.
AssemblyAI STT 3.7 Realtime Agent places highest on the Pipecat Dataset, 2nd of 24 at 1.68%. Its weakest placing is the Ocular Dataset, 12th of 24 at 3.81%.
AssemblyAI STT 3.7 Realtime Agent transcribes the 1,000 public Pipecat clips at 1.68% word error rate and the 300 licensed Ocular turns at 3.81%. Word error rate counts incorrect, missing, and extra words against the reference transcript, so lower is better.
First text arrives after 493ms at the median, and the transcript is final 47ms after the speaker stops. Both are measured on the same 206 streamed turns as every other model, and the page shows P90 and P95 as well, which is what a caller meets on a bad turn.
Every model runs the same two datasets under the same conditions: 1,000 public clips from Pipecat's STT benchmark dataset for accuracy at scale, and 300 turns of licensed Ocular conversations for real speech. Latency comes from 206 streamed turns, using each provider's own supported finalization contract. AssemblyAI STT 3.7 Realtime Agent returned usable text for 1,000 of the public clips.
Cekura runs these simulated calls on your agent and scores every one.