Generates two Gemini candidates (primary plus an automatically chosen secondary model) so the judge can cross-check disagreements. Use when accuracy matters more than latency or cost.
Runs a primary Gemini candidate alongside a Parakeet ASR candidate so the judge can cross-check timing-sensitive transcripts. Use when word-level timestamp precision matters or Gemini timestamps alone are unreliable.
Generates a single Gemini transcription candidate. Use as the fastest, lowest-cost strategy when one strong model pass is sufficient.
Transcribes interviews and Q&A sessions faithfully with speaker turns exactly as spoken. Use when expected_format is "interview" or the audio is a one-on-one interview, journalist Q&A, or research session.
Transcribes lectures, talks, and presentations faithfully, surfacing main points only when clearly spoken. Use when expected_format is "lecture" or the audio is a class, keynote, seminar, or single-presenter talk.
Transcribes legal proceedings, depositions, and hearings preserving exact wording, legal terminology, and attribution. Use when expected_format is "legal" or the audio is a deposition, court hearing, or legal consultation.
Transcribes clinical audio (consults, dictations, rounds) preserving exact medical terminology, drug names, dosages, and speaker attribution without adding interpretation. Use when expected_format is "medical" or the audio is a doctor-patient encounter, clinical dictation, or discusses medication, diagnosis, or pharmacology.
Transcribes multi-speaker meetings and stand-ups with clear speaker turns, preserving action items only when explicitly spoken. Use when expected_format is "meeting" or the audio is a team meeting, stand-up, or status call.