CekuraBench
Book a demo

GPT-4o Transcribe

Every modelOpenAI

Accuracy vs. speed

Lower and further left is better · Nemotron 3.5 Transcribe doesn't support streaming, so it is not shown

Best quadrantBest picks
Final-text delay P50 (ms)
Overall WER (%)

GPT-4o Transcribe is 16th of 19 on the Pipecat Dataset and 18th of 18 on final-text delay.

6.85% Overall WER (19th) · 3.90% Pipecat WER (16th) · 12.45% Ocular WER (19th) · 652ms Final text (18th)Likely range, 95% confidence: Pipecat WER 3.43 to 4.42%, Ocular WER 7.23 to 19.58%Published price $0.36 per hour of streamed audio. OpenAI's estimated cost per minute; billed per token. Source, checked Oct 8, 2026.

Last updated Methodology by Dileep Chagam

GPT-4o Transcribe against every other model

Lower is better on every measure. Hover a column for that model's full numbers.

  1. 1.8%
    Universal 3.6 Pro
  2. 1.9%
    Nemotron 3.5 Transcribe
  3. 1.9%
    Transcribe Stream 1
  4. 1.9%
    Universal 3.5 Pro
  5. 2.0%
    Chirp 3
  6. 2.1%
    Reson8
  7. 2.2%
    Realtime Whisper
  8. 2.4%
    Ink 2
  9. 2.5%
    Linden
  10. 2.5%
    STT-RT v5
  11. 2.6%
    Chirp 2
  12. 2.7%
    STT-1
  13. 3.4%
    4o Mini Transcribe
  14. 3.5%
    Nova-3
  15. 3.7%
    Pulse
  16. 3.9%
    4o Transcribe
  17. 3.9%
    Flux English
  18. 5.0%
    Flux Multilingual
  19. 6.5%
    Gradium
Glossary: WER, TTFT, TTFS, P50
WER · Word error rate
Share of words the model got wrong: replaced, missing or added. Lower is better.
Pipecat Dataset
1,000 short public clips from Pipecat's STT benchmark dataset.
Ocular Dataset
8 recordings from 4 real phone conversations, licensed from Ocular.
Likely range · 95% range
Where the model's true score probably falls, given how much audio was tested. Overlapping ranges are too close to call.
TTFT · Time to first text
How long after the audio starts streaming until the first words appear.
TTFS · Final-text delay
How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.
P50 / P90 / P95 · Typical, slow and slowest turns
Half of turns are faster than P50. Only 1 in 10 is slower than P90, and 1 in 20 slower than P95.

Where GPT-4o Transcribe struggled

Dropped words are its biggest source of error on both real calls and clean clips. Our guide covers testing OpenAI Realtime API voice agents.

86%

of its errors on the Ocular calls are words it left out.

Median across all 19 models: 49%

Mix of errors

Real calls (Ocular Dataset)

Clean clips (Pipecat Dataset)

TTFT and TTFS at P50, P90 and P95

Timed on the same 206 streamed Ocular turns as every other model. P90 and P95 show what a caller meets on a slow turn.

TTFS, final text

How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.

20ms100ms500ms2s
Typical, P50
652ms
Slowest of 18
Slow, P90
1.73s
Slowest of 18
Slowest, P95
1.97s
Slowest of 18

TTFT, first text

How long after the audio starts streaming until the first words appear.

Sends no text until the speaker stops, so first-text timing does not apply.

Typical, P50
—
Slow, P90
—
Slowest, P95
—

Every other model's typical turn, on the same axis.

1,000 Pipecat clips · 8 Ocular recordings · 206 turns timed · 0 failed requests

GPT-4o Transcribe head to head

Every comparisonAll 19 modelsBenchmark source

GPT-4o Transcribe FAQ

  • GPT-4o Transcribe places highest on the Pipecat Dataset, 16th of 19 at 3.90%. Its weakest placing is final-text delay, 18th of 18 at 652ms.

  • GPT-4o Transcribe transcribes the 1,000 public Pipecat clips at 3.90% word error rate and the eight licensed Ocular recordings at 12.45%. Word error rate counts incorrect, missing, and extra words against the reference transcript, so lower is better.

  • GPT-4o Transcribe returns no text at all while the speaker is talking, so a measured time to first text would track how long each turn happened to be rather than anything about the model. Its final-text delay is still comparable and is reported in full.

  • 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 eight licensed conversation recordings from Ocular for real speech. Latency comes from 206 streamed turns, using each provider's own supported finalization contract. GPT-4o Transcribe returned usable text for 1,000 of the public clips.

Building on OpenAI? Test your agent the same way.

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

Book a demoStart free