Environment details

• Programming language: TypeScript / JavaScript
• OS: macOS 15.x (Sequoia)
• Language runtime version: Node.js 22.16.0
• Package version: @google/genai 1.44.0

Steps to reproduce

Model: gemini-2.5-flash-native-audio-preview-12-2025
API: Live API (WebSocket, v1alpha), ephemeral token auth

1. Connect to the Live API with both inputAudioTranscription: {} and outputAudioTranscription: {} enabled in the session config:

   const session = await ai.live.connect({
     model: "gemini-2.5-flash-native-audio-preview-12-2025",
     config: {
       responseModalities: [Modality.AUDIO],
       inputAudioTranscription: {},
       outputAudioTranscription: {},
     },
     callbacks: {
       onmessage: (message: LiveServerMessage) => {
         console.log(message.serverContent?.inputTranscription);
         console.log(message.serverContent?.outputTranscription);
       }
     }
   });

2. Speak several complete sentences, then wait for the model to respond with a complete turn.
3. Observe the inputTranscription and outputTranscription objects logged from serverContent.

Expected behavior

Per the SDK's TypeScript types, and the documented pattern for streaming transcription, the final transcription message for each speaker turn should include finished: true, signaling that the turn is complete and the accumulated text can be finalized.

This is the natural mechanism for knowing when to flush a transcript buffer.

Referenced SDK Patterns

• Live Server Content Interface
• Audio Transcription Interface

Actual behavior

The finished field is never present on either inputTranscription or outputTranscription messages across an entire conversation.

Each transcription message contains only a text field with a fragment. There is no per-transcription signal indicating that a turn has ended.

The only reliable turn-boundary signal is message.serverContent?.turnComplete === true, which arrives on a separate message with no transcription payload. It also fires only at the end of the model's turn, not the user's.

Workaround in use

• Accumulate inputTranscription.text fragments into a buffer.
• Flush the user buffer when the first outputTranscription.text fragment arrives, treating that as an implicit signal that the user has finished speaking.
• Flush the model buffer when turnComplete: true fires.

This works, but it requires developers to reverse-engineer the signal sequence instead of relying on the documented finished flag.