AssemblyAI STT 3.7 Realtime Agent vs GPT-4o Transcribe

Compare

Accuracy and speed

Left gets fewer words wrong. Lower has the final text ready sooner.

Accurate and fast0ms200ms400ms600ms800ms0%2%4%6%Words wrong on clean audio (Pipecat WER)
  • STT 3.7 Realtime Agent1.68% · 47ms
  • 4o Transcribe3.90% · 652ms

AssemblyAI STT 3.7 Realtime Agent 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

Who is better at what

STT 3.7 Realtime Agent4o Transcribe
  • Overall WER2.41%6.85%
  • Pipecat Dataset WER1.68%3.90%
  • Ocular Dataset WER3.81%12.45%
  • Time to first text493msn/a
  • Final-text delay47ms652ms

Lower is better on every measure. Too close to call means the likely ranges overlap.

Clip by clip

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.

STT 3.7 Realtime Agent4o Transcribe
  • ELO rating10241001–1039 · rank 1–16873834–934 · rank 22–23
  • Expected win rate70%30%
  • Ocular turns WER · 110 draws of 19480 won4 won
  • Pipecat WER · 838 draws of 999130 won31 won
  • Semantic WER · 160 draws of 19432 won2 won
  • Phone line · 101 draws of 19893 won4 won
  • Call-center noise · 78 draws of 197115 won4 won
  • Babble noise · 68 draws of 197115 won14 won
  • TV noise · 99 draws of 19870 won29 won
  • Reverb · 97 draws of 19897 won4 won
  • Short replies · 41 draws of 6928 won0 won
  • Entities · 45 draws of 483 won0 won

Short replies are decided on the error count and entities on exact capture, with no draw margin. How the rating is worked out

STT 3.7 Realtime Agent vs 4o Transcribe FAQ

  • AssemblyAI STT 3.7 Realtime Agent is more accurate on both datasets. On the 1,000 public Pipecat clips, AssemblyAI STT 3.7 Realtime Agent scores 2.41% and GPT-4o Transcribe 6.85%, so AssemblyAI STT 3.7 Realtime Agent makes fewer errors by 4.44 points. On the 300 licensed Ocular turns, AssemblyAI STT 3.7 Realtime Agent scores 1.68% and GPT-4o Transcribe 3.90%, so AssemblyAI STT 3.7 Realtime Agent makes fewer errors by 2.23 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.

  • AssemblyAI STT 3.7 Realtime Agent 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 AssemblyAI STT 3.7 Realtime Agent. GPT-4o Transcribe returns no text until the speaker stops, which matters if your agent reacts to partial transcripts.

Test your voice agent the same way.

Cekura runs these simulated calls on your agent and scores every one.