Transcription model that identifies who is speaking when.
GPT-4o Transcribe Diarize is OpenAI's speech-to-text model with built-in speaker diarization, first reported in ~October 2025. Built on the same GPT-4o multimodal architecture as the base gpt-4o-transcribe model, it adds native capability to identify 'who spoke when' — returning speaker labels ("A:", "B:", etc.) with segment-level timestamps. This makes it a direct alternative to dedicated diarization pipelines that previously required piping a separate STT model through a clustering backend.
The model uses the same /v1/audio/transcriptions endpoint and is priced identically to the base transcription model (~$0.0066/min). A key differentiator is the diarized_json response_format, which returns structured segments rather than flat text. Unlike the base model, gpt-4o-transcribe-diarize does not support prompting or the Realtime API, and has an audio chunk limit of 1,400 seconds. Speaker labels are auto-assigned as A/B/C unless explicit speaker references are provided. Per OpenAI's Pete Bakkum, WER is 'roughly comparable' to the base gpt-4o-transcribe — default to the base model if you don't need diarization.
| Type | Rate |
|---|---|
| Audio | $0.0071 / minute of audio |
| Input | $2.70 / 1M tokens |
| Output | $10.80 / 1M tokens |
| Audio input | $6.48 / 1M tokens |
| Estimated | $0.0065 / minute |
request is price per minute (official: same $0.006/min as gpt-4o-transcribe; the earlier +10% diarization surcharge does not exist on OpenAI's rate card)
| Release date | Not publicly disclosed by OpenAI; Azure Foundry lists the model as generally availableai.azure.com ↗ | unknown |
| Architecture | GPT-4o-based speech-to-text model with diarization support for speaker-identified transcriptionai.azure.com ↗ | verified |
| Endpoint | /v1/audio/transcriptions (OpenAI-compatible SDK)developers.openai.com ↗ | verified |
| Language recognition | Improved language recognition for speech-to-text; exact supported-language count for the diarize variant is not publicly disclosedai.azure.com ↗ | unknown |
| Max file size | 25 MB file upload limit for the Transcriptions APIdevelopers.openai.com ↗ | verified |
| Response format | diarized_json returns speaker-labeled segments with start/end timestamps, plus the combined transcriptdevelopers.openai.com ↗ | verified |
| Speaker labeling | Known speaker names are used when references are provided; otherwise speakers are labeled sequentially as A, B, ...developers.openai.com ↗ | verified |
| Supported audio formats | mp3, mp4, mpeg, mpga, m4a, wav, webmdevelopers.openai.com ↗ | verified |
| License | Closed-source / proprietary (API-only access; no open-weight release)openai.com ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
AA-WER (base gpt-4o-transcribe, combined rank) Independent STT benchmark by Artificial Analysis for the base gpt-4o-transcribe. The diarize variant is expected to have comparable WER per OpenAI. No separate AA-WER score exists for gpt-4o-transcribe-diarize as of May 2026. | — | Not publicly disclosed | — |
FLEURS (multilingual WER — base gpt-4o-transcribe) OpenAI reports consistent WER improvements over Whisper on FLEURS (100+ languages) for the base gpt-4o-transcribe architecture. The diarize variant shares the same multilingual backbone. Specific aggregate scores not publicly available. | Improved WER over Whisper v2/v3 across all language groups | openai.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/audio/transcriptions -H "Authorization: Bearer $THE_ROUTER_API_KEY" -F "model=openai/gpt-4o-transcribe-diarize" -F "file=@speech.mp3"Transcribe a multi-speaker audio file and receive speaker-labeled segments using the diarized_json response format.
curl https://api.therouter.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-F "file=@meeting.mp3" \
-F "model=openai/gpt-4o-transcribe-diarize" \
-F "response_format=diarized_json"OpenAI rolled out updated snapshots for gpt-4o-transcribe (2025-12-15) with improved accuracy on Common Voice and FLEURS benchmarks. The same backbone powers the diarize variant. The legacy whisper-1 and initial 2025-03-20 gpt-4o-transcribe snapshots are expected to retire by June 2026, making the diarize model one of the remaining transcription options.
OpenAI silently added the gpt-4o-transcribe-diarize model to the /v1/audio/transcriptions endpoint, offering native speaker diarization on top of the GPT-4o Transcribe backbone. The model uses a diarized_json response format returning speaker-labeled segments with timestamps. Early reports highlight the 1,400-second chunk limit and absence of Realtime API support as key constraints.
The diarize variant adds native speaker identification — it returns structured segments with speaker labels (A:, B:, etc.) and start/end timestamps via the diarized_json response format. The base gpt-4o-transcribe returns plain text without speaker attribution. Caveats: the diarize model does not support the prompt parameter, is not available over the Realtime API, and has a 1,400-second audio chunk limit. WER is 'roughly comparable' per OpenAI, so default to the base model if you don't need diarization.
Speaker labels default to A:, B:, C:, etc. You can provide audio references for speaker attribution (linking segments to known voice profiles), but the model does not learn names permanently — it's per-request attribution, not persistent speaker recognition. Unlike Google Gemini, gpt-4o-transcribe-diarize does not attempt to identify speakers by known identity.
No. gpt-4o-transcribe-diarize is only available on the /v1/audio/transcriptions endpoint for batch/file-based transcription. It does not support the Realtime API or streaming audio input. For real-time transcription, use gpt-4o-mini-transcribe (non-diarized) via the Realtime API or consider Deepgram and AssemblyAI for streaming diarization.
No. This is a key difference from the base gpt-4o-transcribe model. Per OpenAI's speech-to-text guide, prompting is not currently available for gpt-4o-transcribe-diarize. If you need context biasing for domain-specific terminology, use the base gpt-4o-transcribe model instead.
Each request is limited to ~1,400 seconds (~23 minutes) of audio. This is shorter than the base gpt-4o-transcribe's 25 MB / ~25-minute limit. For longer recordings, split into chunks and optionally re-assemble the diarized segments on the client side.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | ai.azure.com ↗ | 2026-08-10 | unknown |
| Architecture | ai.azure.com ↗ | 2026-08-10 | verified |
| Endpoint | developers.openai.com ↗ | 2026-05-28 | verified |
| Language recognition | ai.azure.com ↗ | 2026-08-10 | unknown |
| Max file size | developers.openai.com ↗ | 2026-08-10 | verified |
| Response format | developers.openai.com ↗ | 2026-08-10 | verified |
| Speaker labeling | developers.openai.com ↗ | 2026-08-10 | verified |
| Supported audio formats | developers.openai.com ↗ | 2026-08-10 | verified |
| License | openai.com ↗ | 2026-08-10 | verified |
| AA-WER (base gpt-4o-transcribe, combined rank) | openai.com ↗ | 2026-08-10 | unknown |
| FLEURS (multilingual WER — base gpt-4o-transcribe) | openai.com ↗ | 2026-08-10 | verified |
| OpenAI ships December 2025 snapshot for gpt-4o-transcribe with WER improvements | developers.openai.com ↗ | 2026-05-28 | verified |
| GPT-4o Transcribe Diarize becomes available on OpenAI API | ndurner.github.io ↗ | 2026-05-28 | verified |
| How does gpt-4o-transcribe-diarize differ from the base gpt-4o-transcribe? | ndurner.github.io ↗ | 2026-05-28 | to verify |
| Can I get custom speaker names instead of A/B/C? | learn.microsoft.com ↗ | 2026-05-28 | to verify |
| Is gpt-4o-transcribe-diarize available over the Realtime API? | community.openai.com ↗ | 2026-05-28 | to verify |
| Does gpt-4o-transcribe-diarize support the prompt parameter for context biasing? | platform.openai.com ↗ | 2026-05-28 | to verify |
| What's the audio chunk limit for gpt-4o-transcribe-diarize? | ndurner.github.io ↗ | 2026-05-28 | to verify |