---
title: "Voice Input"
id: "1744"
type: "page"
slug: "speech"
published_at: "2026-09-20T20:56:09+00:00"
modified_at: "2026-09-21T01:36:28+00:00"
url: "https://xedant.com/agents/research/docs/speech"
markdown_url: "https://xedant.com/agents/research/docs/speech.md"
excerpt: "Voice input is dictating tasks by voice straight into the agent’s chat. Press the microphone…"
---

# Voice Input

[https://xedant.com/agents/research/docs/speech.md](https://xedant.com/agents/research/docs/speech.md)

**Voice input is dictating tasks by voice straight into the agent’s chat.** Press the microphone button in the input field and say the task out loud — the recognized speech appears as text in the field, and the message goes to the agent as usual. Typing out a long task is no longer necessary.

## What it is and why

Long tasks are comfortable to dictate — especially on a phone or tablet, where typing is awkward. Say it, and the text is already in the input field; all that remains is to send it. No more “figuring out how to phrase this and typing for five minutes”: say it in your own words, and the agent has its task.

## How to turn it on

Nothing needs configuring: the microphone button is already in the message input of the Research Agent interface. It appears when speech recognition is enabled in the Xedant Agent settings — Research Agent uses its capabilities and has no speech settings of its own. If there is no microphone button, enable speech recognition in the agent’s settings.

## How it works

It is simple: press the microphone button (“Start voice recording”) — recording starts and the button pulses. Say the task, press again (“Stop recording”) — and the recognized text is already in the message field, right where you were typing. Send it as an ordinary message — with the send button or by pressing Enter. The text can be corrected before sending.

There is no need to pause after each phrase. The server-side recognition methods send the accumulated fragment about every five seconds and insert the text as it arrives. Browser recognition shows a live preview of the text above the input field while you speak, and inserts finished phrases at once.

## Recognition engines

The recognition method is chosen in the agent’s settings, and Research Agent adapts to it on its own. Two families work:

- **Browser** — the speech recognition built into the browser (the Web Speech API). It recognizes right in the browser, free and with no keys; it works in Chrome, Edge and Safari. The audio is processed by the browser’s cloud service, so the interface needs an internet connection;
- **Server** — local Whisper on the agent’s server, or cloud services (OpenAI, Groq, OpenRouter, Deepgram, ElevenLabs, and your own OpenAI-compatible service). Research Agent records the voice from the microphone and passes the audio to the agent, which recognizes it with its own engine. Which server engine exactly is configured does not matter to the interface: the model, language and keys are all resolved on the agent’s side.

## What can go wrong

The interface reports every problem with clear hints:

- the first run of local recognition on the agent’s server takes time: the engine installs ffmpeg and downloads the model, and the microphone button shows the hints “installing ffmpeg” and “loading model”. This is one-time preparation — after it, recognition starts immediately;
- the browser does not support speech recognition — Firefox, for example, has no built-in recognition; in that case pick a server engine in the agent’s settings;
- no access to the microphone — allow access in the browser settings; if the microphone is not detected, check the device settings;
- the browser engine needs an internet connection;
- if the recording turned out empty or too quiet, the hint “Check your microphone or speak louder” appears.

## Where it is available

Voice input works in any agent chat in the Research Agent interface — the microphone button sits in the message input. The same button is in every prompt queue row as well: a task can be dictated and added to the queue without typing (see “[Prompt Queue](/agents/research/docs/prompts-queue)
“).

Two clarifications. The recognized text always goes to the agent as an ordinary message — there is no separate voice channel. And Research Agent itself does not recognize speech: it either hands it to the browser, or records it and passes it to the agent — which is why voice input needs a configured, working Xedant Agent.

[← Back to documentation](/agents/research/docs)