CekuraBench
Book a demo

GPT Realtime Whisper

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 Realtime Whisper is 4th of 19 on both datasets together and 15th of 18 on final-text delay.

2.63% Overall WER (4th) · 2.16% Pipecat WER (7th) · 3.53% Ocular WER (6th) · 1.38s First text (11th) · 543ms Final text (15th)Likely range, 95% confidence: Pipecat WER 1.92 to 2.41%, Ocular WER 2.58 to 4.51%Published price $1.02 per hour of streamed audio. OpenAI's estimated cost per minute for live transcription. Source, checked Oct 8, 2026.

Last updated Methodology by Dileep Chagam

GPT Realtime Whisper against every other model

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

  1. 2.2%
    Transcribe Stream 1
  2. 2.3%
    Nemotron 3.5 Transcribe
  3. 2.4%
    Reson8
  4. 2.6%
    Ink 2
  5. 2.6%
    Realtime Whisper
  6. 2.7%
    Universal 3.6 Pro
  7. 2.7%
    Universal 3.5 Pro
  8. 2.8%
    Chirp 3
  9. 3.0%
    Linden
  10. 3.0%
    Chirp 2
  11. 3.0%
    STT-RT v5
  12. 3.6%
    Pulse
  13. 3.7%
    4o Mini Transcribe
  14. 3.8%
    Nova-3
  15. 4.0%
    Flux English
  16. 4.8%
    Flux Multilingual
  17. 5.4%
    STT-1
  18. 5.8%
    Gradium
  19. 6.9%
    4o Transcribe
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 Realtime Whisper struggled

On real calls, dropped words are its biggest source of error. On clean clips, wrong words lead. Our guide covers testing OpenAI Realtime API voice agents.

49%

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
543ms
15th fastest of 18
Slow, P90
647ms
14th fastest of 18
Slowest, P95
669ms
12th fastest of 18

TTFT, first text

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

200ms1s5s20s
Typical, P50
1.38s
11th fastest of 16
Slow, P90
1.84s
9th fastest of 16
Slowest, P95
2.86s
9th fastest of 16

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

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

GPT Realtime Whisper head to head

Every comparisonAll 19 modelsBenchmark source

GPT Realtime Whisper FAQ

  • GPT Realtime Whisper places highest on both datasets together, 4th of 19 at 2.63%. Its weakest placing is final-text delay, 15th of 18 at 543ms.

  • GPT Realtime Whisper transcribes the 1,000 public Pipecat clips at 2.16% word error rate and the eight licensed Ocular recordings at 3.53%. Word error rate counts incorrect, missing, and extra words against the reference transcript, so lower is better.

  • First text arrives after 1.38s at the median, and the transcript is final 543ms 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 eight licensed conversation recordings from Ocular for real speech. Latency comes from 206 streamed turns, using each provider's own supported finalization contract. GPT Realtime Whisper 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