Live API challenges with Gemini 1.5 Flash vs. Pro require careful model selection.

Real-time transcription and speaker identification are incompatible in current implementations. While live transcription provides streaming text, it lacks speaker diarization—only batch processing via the Interactions API offers this feature.

Key differences between live and batch models:

• gemini-1.5-flash-transcribe-live:

Uses WebSocket for streaming and supports real-time subtitles.
No speaker diarization, no word-level timestamps.
Max session duration: 10 minutes.
interimInputTranscription enables live follow-along transcription.

• gemini-1.5-pro-transcribe:

Uses HTTP requests for post-meeting processing.
Supports speaker diarization (up to 8 speakers).
Provides detailed word-level timing.
Max duration: 1 hour (30 minutes with diarization).
Better suited for formal documentation.

For live translation, the gemini-1.5-flash-transcribe-live model is required. Configuration for real-time streams follows this structure:

{
  "setup": {
    "model": "models/gemini-1.5-flash-transcribe-live",
    "generationConfig": { "responseModalities": ["TEXT"] },
    "inputAudioTranscription": {
      "languageCodes": [],
      "mode": "SMART"
    },
    "realtimeInputConfig": {
      "automaticActivityDetection": { "disabled": false }
    }
  }
}

Language detection works automatically with an empty languageCodes array, but specifying ["en-US"] reduces latency. "SMART" mode filters filler words like "uh" or "um," improving readability. The customVocabulary feature supports up to 1,000 entries for specialized terminology.