Speech to Text Without AI: Dictation With No AI Rewrite

Dictation with no AI rewrite is available, but "speech to text without AI" needs a narrower meaning than most people give it. Every modern speech recognizer is machine learning, so no current tool is free of AI in the strictest sense. What you can avoid is three other things: a second model that rewrites what you said, a cloud service that processes your audio, and stored recordings that may be used to improve a product. Several built-in tools type the words they recognize, and where a tool adds an AI pass, its keyboard or dictation settings usually show how to turn it off.

Every modern recognizer uses machine learning, but only a second model rewrites your words

Before about 2012, most recognizers combined hidden Markov models, which handle the timing of speech, with Gaussian mixture models, which score how well each model state fits a short slice of sound ¹. In 2012, research groups reported that deep neural networks did better ². On a Switchboard telephone task, a neural network cut the word error rate of a GMM-HMM baseline from 27.4 to 18.5 percent ². Neural networks replaced the older statistics in the acoustic part of the recognizer.

Whisper, the open model released by OpenAI, is a clear example of a modern recognizer. It was trained on 680,000 hours of multilingual and multitask audio with transcripts ³. A recognizer produces its words as output, so the word "generative" is a poor test on its own. The useful question is whether a second model edits the text after the recognizer has finished.

Table 1 sorts the four things people mean by "without AI" and shows which ones you can avoid.

Pre-neural recognizers typed only with pauses and trained speakers

The older systems did work, but in narrow conditions. DragonDictate, introduced in 1990, cost $9,000 as a system of software and a speech recognition board ⁴. It recognized discrete utterances, which the 1990 report defines as words with at least a quarter-second pause between them ⁵. IBM's Tangora, a research system meant to act as a voice-activated typewriter, had to be trained by each individual user ⁶.

Fluent conversation was much harder. In a 1997 review, an HMM recognizer had a word error rate of 43 percent on the spontaneous telephone conversations of the Switchboard corpus ⁷. Neural systems have since beaten the HMM-GMM systems on the same Switchboard benchmark ², so going back to the older recognizers would mean accepting far more errors. "Without AI" is a choice about what happens after recognition and where the processing runs.

Verbatim dictation gives the recognizer's words, and that is not the same as every word you said

A verbatim setting returns the recognizer's text with no second pass that removes fillers, fixes grammar or tightens sentences. Fillers such as "uh" and "um" make a simple test: if they are missing from the text, either a later pass removed them or the recognizer never produced them. The recognizer can drop them on its own. In a 2025 study of dementia speech, Whisper tended to omit filler words such as uh and um from its transcripts ⁸. So a raw transcript from a Whisper-based tool may already be cleaner than your speech, and you will not see the fillers you said.

Verbatim also does not mean error-free. In one 2024 study, roughly 1 percent of Whisper transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ⁹. That study tested Whisper as it stood in 2023 ¹⁰, so the figure describes that version of the model. Read a verbatim transcript before you send it, because a faithful pipeline can still produce text you did not say.

Table 2 separates the changes that happen after recognition, and whether each one changes your words or only adds marks.

I did not verify how any specific third-party app removes fillers, so the last row describes what to look for, not what a product does.

Built-in dictation types recognized text, and the AI pass Apple documents is new in iOS 27

The standard dictation features I checked describe recognition and punctuation, not a rewrite. The AI pass in Apple's documentation is the iOS 27 feature. On a Mac, Apple says that voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers ¹⁵. Table 3 lists the built-in tools and where each one says recognition runs.

The iOS 27 pass is the one most likely to rewrite your words. 9to5Mac reports that the iOS 27 AI dictation runs only on the on-device model, so its quality is the same with or without a data connection ¹⁸. Birchtree reports that the feature is off by default in the beta ¹⁴, and its steps place the Advanced Dictation Preview switch under Settings, General, Keyboard, scrolled to the bottom ¹⁹. Both are secondary reports of a beta, so check the switch on your own device after the public release.

On Windows, voice typing uses online speech recognition ¹⁶. Microsoft's privacy page says that when online speech recognition is turned off, only device-based speech features, such as Narrator, are available ²⁰. Voice typing depends on online recognition, so it is not one of the features left in that state. Dictation in Office has the same requirement: it needs a microphone and a reliable internet connection ¹⁷.

Built-in tools are not automatically the most accurate. In one 2025 test by 9to5Mac, on a single 7-minute-31-second episode, Whisper Large V3 Turbo had a word error rate of 1.5 percent ²¹. Apple's speech API had 10.3 percent ²², and NVIDIA's Parakeet v2 had 12.3 percent ²³.

Treat this as one test. The Apple result came through an open-source command-line tool, not the Dictation menu ²⁴, so it is not a measurement of the menu you click. For how the Mac side works in practice, see how Apple dictation works, and for Win+H see Windows voice typing.

The AI pass in built-in tools can be switched off, and the defaults differ

The steps below use the menu names from the vendor's documentation, and they describe the systems as of this guide's date. If your version differs, the setting may be in another place.

On a Mac, turn off the automatic punctuation:

1. Open System Settings, then Keyboard.
2. In the Dictation section, find Auto-punctuation and switch it off ¹¹.

On an iPhone, turn off the same kind of marks:

1. Open Settings, then tap General, then Keyboard.
2. Turn off Auto-Punctuation ¹².

On an iPhone running the iOS 27 beta, keep the AI pass off unless you want it:

1. Open Settings, then tap General, then Keyboard, and scroll to the bottom ¹⁹.
2. Leave Advanced Dictation Preview switched off, which is its default in the beta ¹⁴. Birchtree describes switching the feature on, and I did not find an Apple statement on turning it off.

Third-party dictation apps vary. Look for a verbatim or raw mode, a switch for AI formatting, and a statement of whether audio leaves your device. Our answer to how to turn off AI rewriting in dictation covers how to check an app. Yap, which we make, is an on-device dictation app for Mac and Windows, so you can check its settings the same way on the download page.

On-device recognition keeps audio off cloud servers, but only where the tool says so

On a Mac, the choice that matters is Share Audio Recordings. When you enable Dictation, choosing it allows Apple to store audio of your Siri and Dictation interactions, and Apple may review a sample of it. Choosing Not Now declines ²⁵. Leave that choice declined if you do not want recordings kept.

On Windows, online speech recognition uses Microsoft cloud-based services for more accurate recognition ²⁶, so the online setting is the one that involves a service. For dictation that works offline on Windows, read the offline speech to text guide for Windows, which covers what each tool does without a connection.

The on-device and cloud choices can change between versions, so recheck the setting after each major update. For a Mac that stays offline, see offline dictation on Mac.

No source I found says dictated text is flagged differently from typed text

AI detectors score the text they receive, not the keyboard or microphone that produced it. I found no study of dictated text and no statement from a detector vendor about speech input, so the honest answer is that the question is untested. What the sources do show is how the scores behave.

Turnitin's chief product officer wrote in May 2023 that in cases where the detector finds less than 20 percent AI writing, there is a higher incidence of false positives ²⁷.

A 2023 Stanford study tested several widely used GPT detectors and reported that they "consistently misclassify non-native English writing samples as AI-generated" ²⁸. That is a warning about the detectors, not about dictation.

So a dictated draft is judged the way any text is judged: by its wording. Verbatim output stays closest to what you said. A draft that an AI pass has rephrased may read more like model output, but that is an inference, and I have no measured result behind it. For the full answer, see whether AI detectors can flag dictated text and the AI text detector glossary entry.

Two college policies focus on rephrasing and authorship, not on dictation

The two college policies I read turn on whether a model rephrased the words, and on whether the student can declare authorship. One states that students may not have an AI tool rephrase entire sentences or paragraphs of their writing ²⁹. Another leaves some tools out of its definition of generative AI, but it still requires students to be able to declare their authorship for those excluded tools ³⁰. Neither passage I quote addresses speech recognition. Check the rules where you study or work, and keep a record of which tools you used when a policy asks for one.

For the question of whether cleaned-up dictation still counts as your writing, see whether AI-cleaned dictation is your own writing. Related pages explain what AI cleanup changes in your words, how AI dictation works, and whether speech to text is AI. For the overall subject, start with voice and AI.

A short routine for dictation with no AI rewrite

1. Choose a tool whose settings say recognition runs on the device for your language, and check that statement against the tool's own documentation.
2. Turn off automatic punctuation on the Mac or iPhone if you want the marks to come from you.
3. Keep iOS 27 Advanced Dictation off unless you want its spelling and capitalization changes.
4. Decline Share Audio Recordings on a Mac if you do not want recordings kept.
5. Read the transcript before you send it. Look for dropped fillers and for text you did not say.
6. Follow your institution's policy on AI tools, and note which tools you used when it asks.

Speech to text without AI is therefore a set of choices, not a single switch. Recognition is machine learning in every modern tool. What you control is whether a second model rewrites the words, whether audio leaves the device, and whether recordings are kept. The built-in tools cover the first two in several cases, and the settings above show where to change them. Dictation and voice typing covers the typing side of the same choices.