Is Speech to Text AI? What Kind of AI Voice Typing Uses

Yes. Speech to text is a form of AI. The engine that turns your voice into words is a model trained on recorded speech, which is machine learning, and it fits the EU's legal definition of an AI system. Plain dictation transcribes: it picks the words you said. It becomes generative AI only when a language model writes new text on top of that transcript, for example by removing filler words or fixing grammar. Some Windows and Apple dictation now does exactly that. This guide shows the evidence product by product, and explains why the answer changes depending on the policy you have to satisfy.

Speech to text is AI because its recognizer is a trained model

A widely cited legal definition supports the answer. The EU AI Act defines an AI system as a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions ¹. A speech recognizer takes audio as its input and infers the most likely words, so it fits that wording.

The technical evidence points the same way. Apple's 2017 article on its Hey Siri detector says the detector uses a Deep Neural Network to turn the acoustic pattern of your voice into probabilities over speech sounds ². Hey Siri is a wake-word detector rather than dictation, but it shows the same kind of model inside an Apple voice feature. OpenAI's Whisper models were trained on 680,000 hours of multilingual and multitask audio with transcripts ³, which is about 77.6 years of continuous listening ⁴.

The engine's behavior comes from that training, not from a list of hand-written rules. That is what machine learning means in plain terms: the system learns its mapping from examples.

Two terms are easy to mix up. AI is the broad category. Machine learning is the method most modern AI uses. A product can also say "AI" for something else entirely: an assistant that acts on your words, or a rewrite step that edits them. The difference between dictation and voice assistants is covered in its own guide. This guide is about the first two meanings.

Recognition used statistical machine learning long before deep learning

Speech recognition was a machine-learning problem before neural networks took over. In 1952, Bell Laboratories built a system that recognized isolated spoken digits from a single speaker ⁵. The statistical era that followed is described by Juang and Rabiner in their history of the technology. In 2005 they wrote that most practical speech recognition systems were still based on the statistical framework and results developed in the 1980s ⁶.

The standard design of that era had two parts. Hinton and colleagues describe it in their 2012 review: hidden Markov models handled the timing of speech, and Gaussian mixture models scored how well each state fit a short window of audio ⁷. In plain terms, the timing model answered "where in a word am I?" and the scoring model answered "how much does this sound match that state?" Each part was built and tuned by people, and the pieces were combined afterward.

Deep networks replaced the scoring step. The same 2012 review says that deep neural networks with many hidden layers, trained with new methods, had been shown to outperform Gaussian mixture models on a variety of speech recognition benchmarks ⁸. On one Switchboard telephone task, the word error rate fell from 27.4 to 18.5 percent ⁹. That is a relative reduction of 32.5 percent ¹⁰.

The gain was large because the network learned the scoring function from data instead of relying on the designers' assumptions about speech. That is also why the change is a shift in kind, not just a faster version of the same method. A statistical system can be retuned by hand, while a neural system improves when it sees more training audio.

The table shows the direction of travel. Figures from different benchmarks are not comparable with one another, so read each row for its own test.

Transcribing is not generating, until a language model rewrites the text

The distinction that matters is what the output consists of. A recognizer transcribes by choosing words that match the audio. Microsoft describes Copilot dictation as transcribing your speech into text in real time ¹². A transcript is a record of what was said, in the words the recognizer chose.

Generation writes words that were not chosen from the audio. A large language model can do this because it produces text instead of only ranking candidates. In the HyPoradise benchmark, a large language model can correct words that are missing from the recognizer's N-best list, the set of its best candidate transcripts, because its generative capability lets it write text rather than only choose among candidates ¹³. Researchers call this generative error correction. Yang and colleagues study whether large language models can act as post-processors for speech recognition, performing rescoring and error correction ¹⁴.

The line is not perfectly clean. A 2025 preprint on dementia speech notes that Whisper tends to omit filler words such as uh and um from its transcripts ¹⁵. A transcriber that drops fillers is already making a small edit. So "transcript" does not always mean a verbatim record.

A rewrite is also a guess. A language model that corrects grammar has to decide what you meant, and a guess can be wrong about a name, a number or a negation. That is the reason vendors ask you to check the output, covered below. The sequence here shows the two steps as they usually appear in an AI dictation app.

In products, the second step is now a named setting. Microsoft says fluid dictation in Windows Voice Access automatically corrects grammar, punctuation and filler words as you speak ¹⁶, and that the experience is powered by on-device small language models ¹⁷. Apple's iOS 27 dictation, with Apple Intelligence, uses an on-device model to improve spelling, punctuation and capitalization, in English on the iPhone models Apple names ¹⁸. For a plain-language version of this step, see the guide to AI cleanup for dictation.

Both vendors still ask you to check the result. Microsoft tells Copilot users to review the transcribed or generated content for accuracy, because AI may occasionally misinterpret speech or produce errors ¹⁹. That warning applies to the transcript and to the rewrite alike.

Marketing makes the question harder. Product pages often use "AI" for every feature in the same list, from the recognizer that transcribes to the assistant that answers questions. When a page says "AI-powered dictation", the useful follow-up is which of the two steps it means. Ask whether the app can dictate with its rewrite switched off, and whether that switch changes the words or only the formatting. A vendor that can answer both questions in writing is describing its product clearly. One that cannot is asking you to accept the label on trust.

Apple, Google and Microsoft document where dictation runs more than which model runs it

The three companies publish different detail, so the comparison below sticks to what each one says. Where a company does not name the model, this guide does not guess.

Google documents its architecture most clearly. In 2019 it announced an end-to-end, all-neural, on-device speech recognizer to power speech input in Gboard ²⁰. "End-to-end" means one network maps audio to text, rather than several separately built parts. The earlier server recognizer used a search graph of almost 2 GB, which Google said could not easily be hosted on a phone, so it needed an online connection ²¹. Google's 2019 on-device model is 80 MB after compression and keeps working when the phone is offline ¹¹.

Apple documents where processing happens. On a Mac, voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers ²². Apple's newer SpeechAnalyzer developer API transcribes entirely on the device, although its language models must first be downloaded ²³. The lesson for a buyer is that "Apple dictation" can mean different pipelines in different OS versions, so check the version you run.

Independent tests give a sense of the spread. In a 2025 test by 9to5Mac of one 7-minute-31-second episode, Apple's speech API had a word error rate of 10.3 percent, against 1.5 percent for Whisper Large V3 Turbo ²⁴. That is one episode, not a ranking of the products.

Research benchmarks show how far neural models have come. The Conformer, a model that combines convolution with transformers for speech, reached word error rates of 2.1 percent on LibriSpeech test-clean and 4.3 percent on test-other without a language model. With a language model, the figures were 1.9 and 3.9 percent ²⁵. The chart compares the two settings.

Microsoft describes its Windows voice typing as online. Its help page says voice typing uses online speech recognition, which is powered by Azure Speech services ²⁶. Its privacy page says voice typing converts speech to text using online speech recognition ²⁷. Microsoft's page on dictating in Office lists a microphone and a reliable internet connection ²⁸.

The newer rewrite step is limited to particular hardware. Fluid dictation is a Copilot+ PC feature, so it is not available on every Windows 11 PC ²⁹. Copilot dictation is described as transcribing in real time ¹², and its review warning mentions generated content ¹⁹, so the page does not settle whether Copilot rewrites your words.

The table shows what each company documents, and what it leaves out. A blank or "not described" cell means the vendor's page did not say, so you should not assume the feature is absent. Check the settings and the vendor's current page before you rely on either conclusion.

Whether dictation counts as AI depends on the policy, not the technology

The technical answer does not settle a policy question, because each policy writes its own definition. The EU AI Act definition is broad, and speech recognition fits it ¹. Some school and testing rules do not use that definition at all. They can focus on what a tool does for the student. Colorado's spring 2025 security form, for example, requires schools that use speech-to-text on state tests to disclose the name of the application, technology or program students will use ³⁰.

A policy written against generative AI may not reach a plain transcriber, while a policy that defines AI broadly does. So the practical test is mechanical:

1. Does the tool only transcribe, or does it rewrite? Look for a named setting such as fluid dictation, or an optional cleanup mode, and check the vendor's page.
2. Where does the audio go? Apple says Mac general dictation is processed on your device ²². Microsoft's Office dictation needs the internet ²⁸. Those answers matter for data rules even when the AI question is settled.
3. Which definition does the policy use? If it says AI, a transcriber counts under the broad definitions. If it says generative AI or names a product, check whether the rewrite step is on.

A clear policy names the mechanism it cares about, for example "transcription without generated text" or "generative rewriting of submitted work". Naming the mechanism avoids arguments about whether a tool is "AI" in general.

A short example shows the idea. A rule could say that automatic speech recognition that only transcribes your own speech is allowed, and that any tool which generates, rewrites or summarizes submitted text must be disclosed. That wording is our illustration, not a model policy, and anyone who owns a real rule should check it against their own institution's terms. The point is that the rule separates the two mechanisms, so no one has to decide whether a microphone counts as AI in order to follow it.

Yap, which we make, is an on-device dictation app for Mac and Windows. Whether you use it or a built-in tool, the same three questions apply. To try built-in dictation first, see how to use dictation on Mac or Windows voice typing.

Four questions this guide answers in more detail

Each question below has its own short answer and the evidence behind it.

• Is speech to text generative AI?: the difference between transcribing and generating.
• Does Microsoft dictation use AI?: Windows voice typing, Office dictation and fluid dictation.
• Does Google voice typing use AI?: the on-device recognizer in Gboard.
• Does voice typing count as using AI?: how policies define AI and what they regulate.

For the full picture, start with the overview of speech to text AI, then the guides on how AI dictation works, AI cleanup for dictation and speech to text without AI. The explainer on how speech recognition works covers the recognizer itself.