Feature request: let Voice choose what to speak vs. show on screen

Summary

Voice conversations should not treat every assistant response as something that must be read aloud word-for-word.

For humans, speech and reading have very different bandwidth characteristics:

• Speech is a fast output modality: speaking is often much faster and more natural than typing when precision is not critical.
• Reading is a fast input modality: for a user's native language, reading structured information is often much faster than listening to it being spoken linearly.
• This difference becomes even more important for tables, numbers, code, long lists, citations, and other visually structured content.
• For a foreign language, listening can be even slower or harder than reading.

Proposed behavior

In Voice mode, the assistant should be able to make an active multimodal presentation decision for each part of a response:

• Speak concise conclusions, explanations, transitions, and conversational content.
• Put tables, charts, code, long lists, precise numbers, citations, and other dense/structured material on screen without reading all of it aloud.
• When useful, say something short such as: “I’ve put the detailed table on screen—take a look there.”
• Allow the user to continue speaking naturally while using the screen as a high-bandwidth visual channel.

In other words, Voice-first should not mean voice-only. The assistant should decide what is best heard and what is best seen.

Why this matters

Current voice UX can become inefficient when the assistant serializes visually dense content into audio and reads everything aloud. This forces the user to consume information through the slower channel even when a much faster visual channel is already available.

A better interaction model would be asymmetric:

1. The user can speak quickly to express intent.
2. The assistant can answer conversationally by voice.
3. Dense or precise information can appear visually at the same time.
4. Voice only calls attention to the visual content instead of narrating every token.

This would make Voice feel more like a true multimodal interface rather than text-to-speech layered on top of a text chat.

Design principle

Use speech for conversation; use the screen for information density.

Ideally this could be adaptive rather than a manual toggle, while still allowing user preferences for “more spoken detail” vs. “more visual detail.”