OpenAIAudio modelActive
GPT-Realtime-Whisper
PopularityRank 385 of 876
34out of 100same as last month
Summary
GPT-Realtime-Whisper is a dedicated streaming model for real-time transcription. It turns audio streams into text and lets developers tune latency and accuracy. Its documented output path is transcript deltas rather than spoken replies. Unlike GPT-Realtime 2.1, OpenAI documents text output and audio input for the realtime transcription workflow.
The documented context window is 16,000 tokens, with a 2,000-token maximum output. Function calling and structured outputs are not supported. OpenAI lists a September 30, 2024 knowledge cutoff. Its narrow specialization is reflected in the audio-input, text-output contract. OpenAI bills it by audio duration rather than text tokens.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 171 published benchmark results
Page 1 of 15
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Artificial Analysis AA-WER Streaming (final transcription) (Show measurement, test conditions, and source)
- Source value
- 4.8854
- Score
- 4.8854
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- AA-WER v2.2 streaming
- Category
- audio
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Direct API streaming evaluation; final-transcription WER across AA-AgentTalk (50%), VoxPopuli (25%), and Earnings22 (25%), about eight hours total. WER is duration-weighted within dataset; Earnings22 chunking is reassembled at parent-call level. AA-WER v2.2 streaming protocol.
Coval rolling-30-day pooled word error rate (gpt-realtime-whisper endpoint) (Show measurement, test conditions, and source)
- Source value
- 4.061691059994591
- Score
- 4.061691059994591
- Metric
- Pooled word error rate across Coval's fixed public speech-recognition audio conditions
- Unit
- %
- Benchmark version
- Coval rolling 30-day snapshot, last measured 2026-10-16, source response fetched 2026-10-16T15:36:03.942140Z
- Category
- audio
- Direction
- Lower is better
- Sample
- 26641
- Source type
- independent-evaluator
- Evaluator
- Coval
- Status
- superseded
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact hosted API model ID gpt-realtime-whisper. Coval describes daily tests on fixed public clean, accented, noisy, reverberant, far-field, clipped, and phone-codec audio. This endpoint runs Whisper transcription inside an OpenAI Realtime session, not the standalone Audio API. The rolling window was last measured 2026-10-16. Stored score is pooled_value=4.061691059994591 from the WER row, not mean_value; Coval's visible 4.1% is rounded to one decimal. The reported WER sample count comes from that row.
- Context
- The page publishes the WER-scored sample count, not total attempted requests or a failure count. No confidence interval is reported. This hosted Realtime-session result is methodologically separate from Artificial Analysis AA-WER and should not be treated as a replacement or direct equivalent.
Pipecat STT Benchmark semantic WER mean (Show measurement, test conditions, and source)
- Source value
- 2.41
- Score
- 2.41
- Metric
- Arithmetic mean semantic word error rate across successfully transcribed samples
- Unit
- %
- Benchmark version
- Pipecat STT Benchmark at commit 94f567299a21f8d617039bc69f137a131f5dedd5, September 2026 semantic-WER judge update
- Category
- audio
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Pipecat STT Benchmark
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. The pinned results table reports semantic WER over 1000 attempted samples from pipecat-ai/smart-turn-data-v3.1-train. WER Mean is the average across samples with a returned transcript; the table reports 100.0% transcript success, but does not give the exact number of WER-scored samples. Semantic scoring by Claude Sonnet 5.5 ignores punctuation, contractions, fillers, and equivalent phrasings. The source does not document a language distribution; its dataset preview shows English transcripts.
- Context
- No exact WER-scored N or model-specific confidence interval is published. The 1000 figure is attempted samples, not the WER denominator. The README says differences below about 0.25 percentage points are within the 95% bootstrap interval among the most accurate services, so do not rank close scores.
Pipecat STT Benchmark semantic pooled WER (Show measurement, test conditions, and source)
- Source value
- 2.3
- Score
- 2.3
- Metric
- Pooled semantic word error rate, total judged errors divided by total reference words
- Unit
- %
- Benchmark version
- Pipecat STT Benchmark at commit 94f567299a21f8d617039bc69f137a131f5dedd5, September 2026 semantic-WER judge update
- Category
- audio
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Pipecat STT Benchmark
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. The pinned results table reports semantic WER over 1000 attempted samples from pipecat-ai/smart-turn-data-v3.1-train. Pooled WER is total judged semantic word errors divided by total reference words, among samples with a returned transcript; the table reports 100.0% transcript success but not the exact scored N. Semantic scoring by Claude Sonnet 5.5 ignores punctuation, contractions, fillers, and equivalent phrasings. The source does not document a language distribution; its dataset preview shows English transcripts.
- Context
- No exact WER-scored N or model-specific confidence interval is published. The 1000 figure is attempted samples, not the WER denominator. The README says differences below about 0.25 percentage points are within the 95% bootstrap interval among the most accurate services, so do not rank close scores.
FLEURS read-speech streaming WER (Speko) (Show measurement, test conditions, and source)
- Source value
- 5.02
- Score
- 5.02
- Metric
- WER · streaming / Read speech
- Unit
- %
- Benchmark version
- Speko measurement, 2026-09-21
- Category
- audio
- Direction
- No direction published
- Sample
- 1541
- Source type
- independent-evaluator
- Evaluator
- Speko Speech-to-Text benchmark
- Status
- active
- Retrieved at
- 2026-10-19
- Methodology
- Independent evaluation. Measured 2026-09-21; Read speech; n=1541 · FLEURS test+dev+train · vendor-direct; streaming-only endpoint.
- Context
- The retained STT page does not state a WER direction or further transcription protocol details; its linked methodology URL redirects to a TTS methodology article.
Speko Speech-to-Text, conversational WER (Show measurement, test conditions, and source)
- Source value
- 16.8155
- Score
- 16.8155
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Speko conversation track; cell measured 2026-10-01; model-level serving revision not published
- Category
- streaming speech-to-text
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Speko Speech-to-Text benchmark
- Status
- active
- Retrieved at
- 2026-10-19
- Methodology
- Independent evaluation. Exact JSON row openai:gpt-realtime-whisper, displayed as GPT Realtime Whisper; streaming-only endpoint; conversation track, 60 speakers and 2 passes, cell measured 2026-10-01. The JSON scalar is 16.8155 and its display value is 16.82%. The same cell carries the literal annotation ‘14.0–19.9% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01’; its interval level, the meaning of EdAcc, and whether 283 is a transcript denominator are not explained on the opened pages, so no sample count or interval fields are assigned.
- Context
- This is a conversation result and is not interchangeable with the already recorded FLEURS read-speech WER. Speko marks the conversation result as warning-toned while its lower FLEURS value is good-toned. The accessible method link did not resolve to a methodology page; interval semantics, denominator and model revision remain unknown.
Coval rolling 30-day word error rate average (last measured 2026-10-19) (Show measurement, test conditions, and source)
- Source value
- 4
- Score
- 4
- Metric
- Average word error rate (WER), 30-day average; lower is better
- Unit
- %
- Category
- audio
- Direction
- Lower is better
- Sample
- 26644
- Source type
- independent-evaluator
- Evaluator
- Coval
- Status
- superseded
- Retrieved at
- 2026-10-20
- Methodology
- Independent evaluation. Coval reports the hosted GPT Realtime Whisper endpoint, which runs Whisper transcription inside an OpenAI Realtime session rather than the standalone Audio API. This is the page's rolling 30-day average last measured 2026-10-19. The average row reports 26,644 samples. This is a dated snapshot of Coval’s rolling series, not an independent rerun. This average WER is distinct from the previously reported pooled-WER observation.
- Context
- Coval's current average WER is distinct from the older stored pooled WER value 4.061691059994591%, measured 2026-10-16. Do not merge, replace, or interpret this rounded mean as pooled WER. It is a newly observed value from the rolling source, not a new independent study.
Cekura STT benchmark, Pipecat Dataset WER (Show measurement, test conditions, and source)
- Source value
- 2.16
- Score
- 2.16
- Metric
- Word error rate (WER) on short public Pipecat STT benchmark clips against reference transcripts; lower is better.
- Unit
- %
- Category
- audio
- Direction
- Lower is better
- Sample
- 1000
- Lower confidence bound
- 1.92
- Upper confidence bound
- 2.41
- Source type
- independent-evaluator
- Evaluator
- Cekura Bench
- Status
- active
- Retrieved at
- 2026-10-22
- Methodology
- Independent evaluation. GPT Realtime Whisper; 1,000 short public clips from Pipecat’s STT benchmark dataset; word error rate against reference transcripts. Cekura uses each provider’s supported finalization contract.
- Context
- Cekura runs this benchmark itself and sells voice-agent testing and monitoring tools. Its disclosure says evaluated vendors may be customers or partners; some scoring rubrics are not public, so not every part can be reproduced exactly. The source does not identify an immutable serving revision or a measurement date specific to this model.
Cekura STT benchmark, Ocular Dataset WER (Show measurement, test conditions, and source)
- Source value
- 3.53
- Score
- 3.53
- Metric
- Word error rate (WER) on licensed Ocular conversation recordings against reference transcripts; lower is better.
- Unit
- %
- Category
- audio
- Direction
- Lower is better
- Sample
- 8
- Lower confidence bound
- 2.58
- Upper confidence bound
- 4.51
- Source type
- independent-evaluator
- Evaluator
- Cekura Bench
- Status
- active
- Retrieved at
- 2026-10-22
- Methodology
- Independent evaluation. GPT Realtime Whisper; eight Ocular-licensed recordings from four real phone conversations; word error rate against reference transcripts. The recordings are not eight separate conversations. Cekura uses each provider’s supported finalization contract.
- Context
- Cekura runs this benchmark itself and sells voice-agent testing and monitoring tools. Its disclosure says evaluated vendors may be customers or partners; some scoring rubrics are not public, so not every part can be reproduced exactly. The source does not identify an immutable serving revision or a measurement date specific to this model.
Ocular AI Converse-STT Version 2 pooled WER (Show measurement, test conditions, and source)
- Source value
- 6.71
- Score
- 6.71
- Metric
- Pooled word error rate (WER) over reference words
- Unit
- %
- Benchmark version
- Converse-STT Version 2.0, last updated 2026-10-13
- Direction
- Lower is better
- Sample
- 4
- Source type
- independent-evaluator
- Evaluator
- Ocular AI Converse-STT Benchmark
- Status
- active
- Retrieved at
- 2026-10-23
- Methodology
- Independent evaluation. Ocular AI's open-source ocular-evals harness evaluated four two-person American English conversations (~60 minutes, eight speaker tracks) delivered as 48 kHz, 24-bit mono FLAC with verbatim, time-aligned transcripts. WER pools substitutions, deletions, and insertions over 12,555 reference words after Whisper English normalization. The API serving revision and additional request settings are not published.
- Context
- The 12,555 reference-word denominator is not the sample count; N is four conversations. No confidence interval or independent repeat count is stated. This is a fixed public conversation set, not a population estimate.
Coval GPT Realtime Whisper time to final segment (Show measurement, test conditions, and source)
- Source value
- 598
- Score
- 598
- Metric
- Mean time to final segment, ms; lower is better
- Unit
- ms
- Benchmark version
- Coval rolling 30-day benchmark page, captured 2026-10-17; daily/rolling service data
- Direction
- Lower is better
- Sample
- 26641
- Source type
- independent-evaluator
- Evaluator
- Coval, GPT Realtime Whisper benchmark
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact Coval route gpt-realtime-whisper, hosted by OpenAI; rolling 30-day real-time service evaluation. The page names the API route but does not disclose its immutable serving revision, client/network setup, or confidence interval.
- Context
- Coval's public page reports rolling service evaluations. The capture date does not identify an immutable provider-side serving revision, complete request parameters, or a fixed evaluation release; no confidence interval is published.
Coval GPT Realtime Whisper time to first token (Show measurement, test conditions, and source)
- Source value
- 1826
- Score
- 1826
- Metric
- Mean time to first token, ms; lower is better
- Unit
- ms
- Benchmark version
- Coval rolling 30-day benchmark page, captured 2026-10-17; daily/rolling service data
- Direction
- Lower is better
- Sample
- 26641
- Source type
- independent-evaluator
- Evaluator
- Coval, GPT Realtime Whisper benchmark
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact Coval route gpt-realtime-whisper, hosted by OpenAI; rolling 30-day real-time service evaluation. The page names the API route but does not disclose its immutable serving revision, client/network setup, or confidence interval.
- Context
- Coval's public page reports rolling service evaluations. The capture date does not identify an immutable provider-side serving revision, complete request parameters, or a fixed evaluation release; no confidence interval is published.
Page 1 of 15
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
0 to 100, higher is more popular
20 of 83 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 6, 2026
GPT-Realtime-WhisperGPT-4o Mini TTS 36Gemini 3.5 Transcribe Live 35GPT-Realtime Translate 35GPT-4o Mini Transcribe 33
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
GPT-Realtime-Whisper compared
- 1GPT-Realtime-Whisper vs. GPT-Live-Transcribe1 shared test
- 2GPT-Realtime-Whisper vs. Flux2 shared tests
- 3GPT-Realtime-Whisper vs. Nova-31 shared test
- 4GPT-Realtime-Whisper vs. Gemini 3.5 Transcribe Live4 shared tests
- 5GPT-Realtime-Whisper vs. Scribe v2 Realtime3 shared tests
- 6GPT-Realtime-Whisper vs. Voxtral Mini Transcribe Realtime19 shared tests
- 7GPT-Realtime-Whisper vs. Cartesia Ink-21 shared test
- 8GPT-Realtime-Whisper vs. Muse Voice Transcribe2 shared tests
- 9GPT-Realtime-Whisper vs. Universal-3.6 Pro Realtime1 shared test
Similar models
More models of the same kind as GPT-Realtime-Whisper, chosen by model line, developer, size, price and release date.
- GPT-Realtime 2.1OpenAI · same developer
- GPT-Live-TranscribeOpenAI · same developer
- GPT-Realtime TranslateOpenAI · same developer
- GPT-TranscribeOpenAI · same developer
- GPT-Realtime 2.1 MiniOpenAI · same developer
- GPT-Live 1OpenAI · same developer
- Gemini 3.5 TranscribeGoogle · popular in this category
- Nova-3Deepgram · popular in this category
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
API
Further details
Prices
- Live transcription price
- $0.017 per audio minuteSource
Sources and data date
Show all 26 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.