Voice, screen readers and hands-free use

A route decision, not an afterthought

For a lot of people the question is not which model is best but which route they can actually use: someone with limited vision, with a motor impairment, with dyslexia, with a repetitive strain injury, or someone whose hands are busy. The tools differ enormously here, and the differences are rarely mentioned in any comparison.

Speaking instead of typing

Two different things get called voice input and they are worth separating.

Dictation turns speech into text in the input box, which you then edit and send. Accurate, cheap, works everywhere, and it is the better option for anything long or precise. Your phone’s built-in dictation is free and often as good as anything in the app.

Voice mode is a spoken conversation: you talk, it talks back, you can interrupt. Genuinely good for thinking aloud and for hands-free use. Worse for anything you need to keep, because the reply is speech and reviewing it means reading a transcript anyway.

Accent handling has improved and remains uneven. Speech recognition trained mostly on a few accents does measurably worse on others, and this is one of the clearest places where the technology works better for some people than others. Test it in your own accent, in your own language, before you build a habit on it.

Free paths are strong. Whisper transcribes dozens of languages and runs on an ordinary laptop; whisper.cpp makes it fast on a CPU. Dictating into any text field and pasting is available to everyone at no cost, without a subscription that includes voice.

Having it read back to you

Text-to-speech is built into every operating system and is free: VoiceOver on Apple devices, Narrator or the Read Aloud feature in Windows, Select-to-Speak on Android, Orca on Linux. For better voices at no cost, Piper runs locally and produces natural speech offline in many languages.

This matters for a use most people miss: having long AI output read to you while you do something else is a very good way to review it. Errors that the eye skims over are noticeable when heard.

Screen readers and chat interfaces

Streaming replies are awkward with a screen reader. Text appears token by token, and depending on how the page announces updates, the reader may repeat the growing answer, interrupt itself, or read nothing until the end.

Products differ, and this is worth testing before committing:

• Does the answer get announced once when complete, rather than continuously?
• Are the controls (model picker, attach, regenerate, stop) reachable and labelled with the keyboard?
• Is the conversation a sensible reading structure, or a wall of unlabelled containers?
• Can you turn streaming off? Some products offer this and it usually improves screen-reader use immediately.

Where the official app is poor, the free front-ends from earlier in this module are worth trying, since they are ordinary web applications and some are more standards-compliant than the commercial products. Where none is workable, the API route with a plain text client is fully accessible by construction.

Hands-free and keyboard-only

Keyboard-only use is mostly a matter of shortcuts: a global hotkey to open the assistant, and a reliable way to send without reaching for a mouse. Desktop apps generally do this better than websites.

For heavier needs, Talon and Dragon drive a computer by voice; the free and open Serenade does something similar for coding. These are general accessibility tools rather than AI tools, and they combine well with any of the routes here.

Reading and writing difficulties

One use deserves naming on its own. For people with dyslexia, these tools are unusually good at a specific job: taking text you have written and fixing spelling, punctuation and structure without changing what you meant. That is a task with a checkable answer, it does not require the model to know anything, and even a small local model does it well.

The same works in reverse. Asking for a long document to be restated in plain sentences, or as a numbered list, makes material accessible that was written for somebody else. Both of these are ordinary chat, need no special product, and are worth more to some readers than every feature comparison in this course.

The point worth making plainly

An assistant you can speak to and be read back by is, for some people, the difference between doing a piece of work independently and needing help with it. That is a larger effect than any difference in model quality this course discusses, and it is a reason to test the route before the model: the best model in an interface you cannot navigate is worth less to you than a middling one in an interface you can.