multimodal audio in vs. transcribing an audio attachment?

Currently, it appears that audio is handled as a binary attachment, and is then transcribed. ref https://github.com/flutter/ai/blob/b1fbc7c8dfc6b1b0dc7c2485b4ea2f3d99befa32/lib/src/views/llm_chat_view/llm_chat_view.dart#L230

For multimodal models such as Gemini, audio as an input is natively supported.

The expectation is that instead of an audio attachment that is transcribed, the audio should be used as the input to the model directly rather than the transcription.

Start with lib/src/views/llm_chat_view/llm_chat_view.dart around line 230 and trace how an audio attachment becomes a transcription. Compare that flow with the multimodal model input path. Done means supported audio is passed directly to the model instead of being transcribed first.

Written by the indexing model from the issue text.