benchrAll capabilities Image, audio and video
Transcribe a recording and label who said what
Speech to text is the easy half. The useful half is a transcript where each line is attributed to a speaker.
DocsOfficially documentedTestNot tested by benchr
OpenAI API
Overview
- Transcribes with gpt-transcribe as the general recommendation, with whisper-1 and the older gpt-4o transcribe models still documented.
- Separates speakers with gpt-4o-transcribe-diarize, which returns a diarized_json response format.
- Returns word and segment timestamps on whisper-1 through timestamp_granularities, which is what makes a transcript navigable.
How it works
- 01Split the recording before uploading. The documented maximum is 25 MB per file, which a long meeting exceeds easily.
- 02Choose deliberately: diarization and word-level timestamps are documented on different models, so pick for the output you need.
- 03Keep the audio in a supported container - mp3, mp4, mpeg, mpga, m4a, wav or webm.
- 04Feed the transcript to a second pass for summarising or extracting actions. Transcription and comprehension are separate jobs.
Limits
- 25 MB per file is a hard ceiling; chunking is your problem, and chunk boundaries can split a sentence and a speaker turn.
- Timestamp granularity and diarization are not offered by the same model, so one request cannot always give you both.
- Speaker labels are identifiers, not identities - the model separates voices, it does not know who they are.
Verification
benchr test runs
benchr has not run this capability. No test has been published.
benchr test runs →
Evidence
Sources
- OpenAI - Speech to text: models, 25 MB limit, formats, diarization and timestamps Tier 1 · September 1, 2026
Not stated by the source
- Word error rate on accented or noisy audio
- Maximum audio duration once the file-size limit is respected
Source-check history
- September 1, 2026DocsOfficially documentedFirst logged. Model names, the 25 MB ceiling, the format list and which model carries diarization read from the guide.