Two Very Different Approaches to Voice

There are two ways to add voice to an AI agent. You can build the entire experience around voice from the start. Or you can build a text-based agent and add a microphone button later. These produce fundamentally different products.

What Voice-Added Looks Like

Take a text-first agent and add speech-to-text. The user speaks, their words get transcribed, and the agent processes the text exactly as if it were typed. Responses come back as text that you read on screen.

This works, technically. But it creates friction everywhere. The agent expects precise, structured input because it was designed for typing. Natural speech is messy - it has filler words, restarts, and implied context. A voice-added agent fights against how people actually talk.

What Voice-Native Looks Like

A voice-native agent assumes you're speaking. The entire interaction model is designed for it. Responses are concise and spoken back to you. Confirmations are quick yes/no exchanges, not paragraphs of text. Ambiguity gets resolved through brief spoken clarification, not long text prompts.

The UX differences cascade through every interaction. Error handling, progress updates, task confirmation - all of it changes when you design for voice first.

Why It Matters for Desktop Agents

Desktop agents are supposed to free your hands. If you're still staring at a screen reading text responses, you haven't actually gained much freedom. Voice-native design lets you issue commands while doing something else entirely - reviewing a document, sketching on paper, or just thinking.

The agent that speaks to you and listens naturally fits into your workflow. The one that transcribes your speech and shows you a text box doesn't.

Building voice-native is harder. You can't just add a microphone icon and call it done. But the result is an agent that actually feels like talking to an assistant, not dictating to a chatbot.

Fazm is an open source macOS AI agent. Open source on GitHub.