What is speech to text in AI?

Speech to text in AI means software that takes spoken audio and writes it down as words. The engine that does this is a model trained on recorded speech with matching transcripts, so its behavior comes from examples rather than from a rulebook written by hand. The input is sound and the output is text. In plain dictation, the words on the page are the words the recognizer picked from your audio.

The input is sound and the output is the words chosen from it

The input to a speech recognizer is a series of acoustic waves ¹. The recognizer turns those waves into the most likely sequence of words. Microsoft describes its speech to text service as providing solutions for converting audio streams into text ². Those streams can come from a live microphone or from a recording that is already finished.

Automatic speech recognition is the technical name for this task. Most people just say speech to text, voice to text or voice typing, and in everyday products the three usually mean the same thing.

It is AI because the recognizer is a trained model

The AI part is how the recognizer is built. OpenAI's Whisper models were trained on 680,000 hours of multilingual and multitask audio with transcripts ³. The model adjusts itself until its output matches the transcripts, and that adjustment is machine learning. Machine learning is the method behind most of the AI people use today, so a speech recognizer built this way is AI in the technical sense.

Training data is not the same as accuracy. A large training set tells you how much the model saw, not how well it will hear your voice, your accent or a noisy room. Accuracy has to be measured on the kind of audio you actually use.

Speech to text has been part of AI since the field began

The idea is older than most products. The 1955 proposal for a summer research project on artificial intelligence said an attempt would be made to find how to make machines use language ⁴. A 2026 draft of a standard textbook on language processing describes understanding spoken language, or at least transcribing it into writing, as one of the earliest goals of computer language processing ⁵.

The methods changed a lot. Earlier systems combined hidden Markov models with separately built scoring parts. By 2012, hybrid systems that pair neural networks with those models had surpassed the older systems in the textbook's review ⁶. That shift helped make modern recognizers good enough for everyday dictation.

The common names describe one task in different settings

People use several names for the same underlying job. The table shows how they differ in practice.

The last row is the one most often confused with the others. An assistant uses speech recognition as its first step, then does something with the words. The guide to voice assistants and dictation explains that split.

It depends on where the audio is processed

The task is the same whether the recognizer runs on your device or on a server, but your data moves differently. Microsoft says voice typing on Windows converts speech to text using online speech recognition ⁷. Apple says that on a Mac, general Dictation is processed on your device and not sent to Siri servers ⁸. Both are speech to text. The privacy answer differs, so check the settings and the current vendor page for the version you run.

A second case is the rewrite. A recognizer that only transcribes is not generative. A language model that edits the transcript is generative, because it writes words that were not chosen from the audio. The question on whether speech to text is generative AI covers that difference.

What to do if you need speech to text with AI

1. Name the job. Decide whether you want the words you said, written as they come, or cleaned-up prose. The first is transcription. The second adds a rewrite step.
2. Find out where the audio goes. Read the privacy or data page for the product, and check whether the feature needs an internet connection.
3. Test on your own audio. Dictate a short passage you know by heart, including your names and terms. Count the wrong words, or use a word error rate check, before you trust a tool.
4. Keep the rewrite off when accuracy matters. If the app has a cleanup switch, compare the output with it on and off, because a rewrite can change a name, number or negation.

Yap, which we make, is an on-device dictation app for Mac and Windows. It is one of several ways to run speech to text on your own machine, and the built-in tools are free and may be enough. For a fair comparison of the options, see the guide to AI dictation.

Where to read more

The short answer to the related question of how speech to text differs from AI is on the question comparing speech to text and AI. For the full set of products and policies, read the guide to whether speech to text is AI, and the overview of speech to text AI that covers the whole subject.