Blog

Voice to text software converts spoken words into written text in real time. Twenty years ago it was a specialist tool that required an hour of voice training and a quiet room. Today it is accurate enough out of the box that most professionals can dictate a full document faster than they can type it.

This guide covers what the technology actually does, what to look for, and where the current options differ. It is written from 28 years of deploying speech recognition across Australian hospitals, courts, and government departments.

How voice to text actually works

Modern speech to text is built on neural networks trained on very large volumes of speech. Rather than matching sounds to words one at a time, the system predicts whole sequences using context. This is why current software can distinguish "their", "there", and "they're" correctly, and why it handles Australian accents far better than systems from even five years ago.

Two consequences matter in practice. First, voice training is largely obsolete: most platforms are accurate from the first sentence. Second, accuracy improves with context, so dictating full sentences produces far better results than speaking word by word. The most common mistake new users make is speaking too carefully.

What "99% accuracy" really means

Vendors quote accuracy figures freely, and they are not meaningless, but they need reading carefully. At 99% accuracy you still get roughly one error every hundred words. Over a two-page letter that is six or seven corrections.

What separates good software from adequate software is usually not the headline figure but three practical things: how well it handles your specialist vocabulary, whether it punctuates automatically, and whether it works inside the application you actually use. A system that is 99% accurate in its own window but cannot dictate into your practice management software is worth less than one at 97% that works everywhere.

Desktop, cloud, or built-in

Built-in dictation (Windows Voice Access, Microsoft 365 dictation, macOS) costs nothing and is genuinely useful for short messages. It has no specialist vocabulary, limited formatting control, and no custom commands. For occasional email it is fine.

Desktop software installs locally and can run without an internet connection once activated. This matters where connectivity is unreliable or where data handling requirements prevent audio leaving the network. It also supports deep customisation: custom vocabulary, text shortcuts, and voice macros that automate repetitive workflows. Dragon Professional 16 is the established option in this category and is sold as a perpetual licence.

Cloud platforms process audio on remote servers, which keeps the local application lightweight and means accuracy improves without you reinstalling anything. A current cloud speech to text platform installs in about two minutes, needs no voice training, and dictates at your cursor in any Windows application. Pricing is typically subscription-based and considerably lower than a perpetual desktop licence. The trade-off is that you need a working internet connection.

Specialist vocabulary is where most projects succeed or fail

General-purpose voice to text handles everyday language well. It does not know "metacarpophalangeal", "subpoena duces tecum", or your suppliers' product codes. If your work involves specialist terminology, general software will frustrate you within a day.

This is why medical and legal editions exist. They ship with domain vocabularies covering tens of thousands of specialist terms. For clinicians in particular, the gap is stark: general software transcribing a clinical note will misrender drug names and anatomy in ways that create real documentation risk. If that is your use case, look specifically at medical dictation software built for Windows EMRs rather than adapting a general tool.

Microphones matter more than people expect

The single most common cause of disappointing accuracy is a poor microphone. Laptop microphones capture room noise, keyboard clatter, and air conditioning along with your voice, and no amount of software cleverness fully recovers from that.

A dedicated USB headset or handheld dictation microphone will usually improve accuracy more than switching software will. Bluetooth earbuds, including AirPods, are generally a poor choice for sustained dictation because they compress audio heavily. If you are evaluating voice to text and getting mediocre results, change the microphone before you change the platform.

How to evaluate options properly

1. Test with your own work. Dictate a real document containing your actual terminology, not a sample paragraph.
2. Test in your real application. Confirm it dictates directly into the software you use daily, not just into a text box.
3. Time the correction, not the dictation. The measure that matters is total minutes to a finished document.
4. Check where your data goes. For clinical, legal, and government work, establish whether audio is stored and where it is processed.
5. Use the trial period properly. Accuracy on day five is more informative than accuracy on day one.

Realistic expectations

Most people reach roughly three times their typing speed within a fortnight. The first few days feel slower, because composing out loud is a different skill from composing at a keyboard, and dictation rewards thinking a full sentence through before speaking it. Users who abandon voice to text almost always do so in the first week, before that adjustment happens.

Set aside the first two days as a learning period, use a proper microphone, and add your specialist terms to the custom vocabulary early. Those three things account for most of the difference between people who find dictation transformative and people who conclude it does not work.