audio model
Gemini 3.5 Transcribe
GoogleThe best Greek in our harness, and the only model here with speaker attribution.
Two models share this name and they are not interchangeable. The Live sibling is the fastest streaming transcription we have measured — 666ms of text still arriving after an English speaker stops, against 929ms for OpenAI’s streaming model. The recorded-audio model is slower but carries speaker attribution and word-level timestamps, which nothing else in our stack offers. For a panel interview that distinction decides which one you want.
Variants
2 ways to run it.
Same family, different quality/cost rung. Picking the wrong one is the most common way teams overpay.
Live
Fastgemini-3.5-transcribe-liveWebSocket streaming. Fastest tail latency we have measured — 666ms EN / 876ms EL.
Live captions · Real-time interview notes
Recorded
Highgemini-3.5-transcribeSlower (2.4s EN / 2.6s EL end-to-end on opus) but adds speaker attribution and word timestamps.
Panel interviews · Anything needing who-said-what
Fit
Two audiences, one model.
The same model is a different proposition depending on what you point it at.
Recruitment fit
The pick for panel interviews and any non-English hiring — speaker attribution and Greek accuracy are both unmatched here.
Best for
- Panel interviews
- Multilingual hiring
- Evidence-linked interview notes
Strengths
- Speaker attribution turns a panel recording into an attributable transcript
- Word-level timestamps let you deep-link a claim back to the moment it was said
- Best measured Greek WER in our harness (0.03)
Watch out for
- Live and Recorded are different model codes — you cannot swap one for the other
- Requires a billed key; the free tier 429s under sustained use
The numbers
Measured, not asserted.
HumanLike evals
Run on our own harness.
- Word error rate (English)
- 0.06Our eval
- Word error rate (Greek)
- 0.03Our eval
- Text still arriving after speaker stops (EN)
- 666 msOur eval
Best Greek of any model in our harness except gpt-transcribe.
Against 929ms for gpt-live-transcribe and 1197ms for the Groq default.
Compare
