audio model
GPT Transcribe
OpenAIThe only model in our harness that does not degrade under background noise.
GPT Transcribe holds the "Smart" tier in HumanLike Coworker, and the reason is counter-intuitive. On clean audio it is the *worst* of the four models we benchmarked — 0.29 WER against mini’s 0.12. Add background noise and the ordering inverts: it lands at 0.26 while mini collapses to 0.42. It is the only model we tested that does not get worse when the room does, and real dictation happens in real rooms.
Variants
3 ways to run it.
Same family, different quality/cost rung. Picking the wrong one is the most common way teams overpay.
Smart
Highgpt-transcribeHighest quality for complex or technical content, and the most noise-tolerant.
Interviews recorded in real rooms · Technical content
Balanced
Balancedgpt-4o-mini-transcribeBest clean-audio accuracy in our harness — and the worst under noise.
Studio-quality audio · Cost-sensitive volume
Live (streaming)
Fastgpt-live-transcribeWebSocket-only. Text streams while the speaker is still talking — ~300ms of text remains after they stop.
Live captioning · Real-time interview notes
Fit
Two audiences, one model.
The same model is a different proposition depending on what you point it at.
Recruitment fit
The right transcription default for interviews, because interviews are recorded in noisy rooms on laptop mics.
Best for
- Interview recordings
- Screening call notes
- Technical role discussions
Strengths
- Only model in our harness that holds accuracy under background noise
- Handles technical vocabulary and proper nouns — job titles, tool names, company names
- Cheap enough at $0.0045/min to transcribe every call
Watch out for
- On clean studio audio the Balanced tier is genuinely more accurate — do not blanket-apply Smart
- Small benchmark samples: re-validate on your own audio before standardising
The numbers
Measured, not asserted.
HumanLike evals
Run on our own harness.
- Word error rate (clean audio)
- 0.29Our eval
- Word error rate (with additive noise)
- 0.26Our eval
- Median latency (9.2s clip)
- 1390 msOur eval
Small sample. Indicative, not definitive. Worst of four models on clean audio.
Best of four. The only model that did not degrade under noise.
Equivalent to gpt-4o-transcribe (1.45s), not an improvement.
Compare
