Gemini Live API: distinguish a spoken response from a completed action

Start with the personal-assistant example

The October 2025 video tests Gemini’s Live API inside Google AI Studio: https://aistudio.google.com. It covers real-time conversation, voice choices, a native-audio function-call sandbox and a personal assistant. The description’s free-access language belongs to that recording, not verified current API or product terms.

The initial demo is at 0:47, AI Studio setup at 2:04 and assistant creation at 3:50. Choose one small job before adding multiple tools. A clear conversational reply is useful, but it is not proof that a calendar entry or email was created.

1. Follow the recorded AI Studio setup and voice settings.
2. Describe one assistant job and the information it needs.
3. Inspect a short conversation before introducing a function call.

Check the actual function result

Function calls are tested at 6:16. The source discusses calendar events and email but explicitly says function calling did not work perfectly. Inspect the requested action, its supplied details and the real result rather than treating a spoken confirmation as completion.

Keep the assistant’s intended task narrow enough to check. For a calendar example, compare the requested event details with the returned tool evidence or saved event; for an email, inspect its recipient and complete content before any intended send.

1. Record the requested action and expected result.
2. Inspect the tool response and destination state.
3. Keep a failed or incomplete action visible before retrying.

Assess language and export separately

Spanish begins at 7:44, Scottish Gaelic at 8:47 and the email function test at 11:17. Compare how well the assistant understands and responds to the chosen language without assuming that fluency establishes correct external actions. The description labels email sending as attempted.

The source also shows exporting to GitHub for work in an IDE and discusses why dedicated voice tools may still be useful. Treat export, conversation quality and reliable tool execution as distinct requirements when deciding whether to develop the prototype further. The record supplies no universal replacement claim for specialised services.

The takeaway

Check the voice interaction and the destination of each requested action independently, preserving the recorded failures.