Speech to Text for AI: Why Spoken Prompts Work Better

Speech is about three times faster than typing on a phone ¹, which makes the detail a model needs cheaper to say. A controlled study put English speech input well ahead of the keyboard. Language models can recover some misheard words and tolerate a lot of random noise, but they are less forgiving of wrong names, code and numbers, and a single chosen typo lowered accuracy on a math benchmark ². This guide is part of our overview of speech to text and AI. It covers the evidence, how to structure a spoken prompt so it survives transcription, and where voice prompting still fails.

Speech is about three times faster than typing on a phone

In a controlled study of short messages on an iPhone, English text entered by speech reached 153 words per minute, against 52 words per minute on the built-in keyboard, 2.93 times faster ^{1 3}. The study counted the finished text, including the time spent correcting mistakes, which is the fair measure for a prompt you will send. It used short phrases in a quiet room, and its authors note that results in everyday settings still need study ⁴.

An online typing test with 168,960 participants found that average typing speed was 51.56 words per minute ^{5 6}. Speech at 153 words per minute ¹ is about three times that average ⁷. The two studies used different people, devices and tasks, so treat the ratio as a rough guide rather than a measure of your own speed. For more on typing rates in general, see average typing speed and speaking vs typing speed.

The speed gap matters most for long text. At the study's rates, a 1,000-word draft takes about 6.5 minutes to speak and 19.2 minutes to type, a saving of about 12.7 minutes ⁸. Conversational telephone speech is faster still, at about 3.8 times the average typing speed in the online test ⁹. The figures suggest that speaking a long request can be quicker than writing it out, once you already know what you want to say.

Speed is not the same as more detail. Speaking may make it easier to add context, but we found no controlled study that measures whether spoken chatbot prompts carry more of it than typed ones. The one large log study of voice search we could find, from 2010, found that longer queries were more likely to be typed than spoken ¹⁰. The evidence for spoken prompts is about speed, and about the fact that speaking lowers the cost of saying the extra detail. It does not show that spoken prompts are richer.

Longer, more specific prompts usually get better answers, up to a point

Longer prompts tend to help when they add information the model needs. A 2025 study of domain-specific tasks found that longer prompts generally improved model performance, while shorter prompts could be detrimental ¹¹. The longer versions in that study added background knowledge, and even with that knowledge the models still struggled with challenging domain-specific tasks ¹². Length helps when it adds facts, constraints and examples. For the direct question, see do longer prompts get better AI answers.

Anthropic's guidance for Claude gives the same practical rule. It asks writers to treat the model as a brilliant but new employee who lacks context, and says that the more precisely you explain what you want, the better the result ¹³.

Where the detail sits also matters. In long-input tests, models were often most accurate when the relevant information came at the start or the end, and accuracy significantly degraded when it sat in the middle ¹⁴. Those tests used long inputs, not short chat prompts, so they do not measure your prompt directly. They still suggest a simple habit: say the task first, then the material, then the constraints that matter most.

Wording and layout matter as well. In few-shot prompts that differed only in formatting, LLaMA-2-13B's accuracy varied by up to 76 accuracy points ¹⁵. That study used open-source models with worked examples, so the size of the effect may not carry over to a chat prompt. It does show that the same request can get a different answer when it is phrased or laid out differently.

Language models recover some misheard words, but chosen errors can still flip answers

A dictated prompt contains fillers, false starts and the recognizer's mistakes, and the evidence treats each differently. Whisper can drop fillers before they reach the text. In a 2025 study of speech from people with dementia, it tended to leave out filler words such as uh and um ¹⁶. Fillers may also mark timing rather than content. Clark and Fox Tree proposed that speakers say uh before a minor delay and um before a longer one ¹⁷. So a prompt full of "um" is still a prompt, and the words that carry the request are the ones around the fillers. For more on what a cleanup step removes, see AI cleanup for dictation.

Misheard words are a different case, and models can repair some of them. In the HyPoradise benchmark, a large language model can correct words that are missing from the recognizer's list of candidates, because it generates text rather than only choosing among the candidates ¹⁸. The benchmark built on more than 334,000 pairs of recognizer hypotheses and correct transcripts ¹⁹. A 2024 study found that giving a model word alternatives from the recognizer, instead of only its top guess, helped bridge the gap to an oracle upper bound ²⁰. These methods are research setups, and a chat user cannot pick them. The practical lesson is that a model can sometimes recover a misheard common word from context. The studies here do not show that it recovers a rare name or a number, so check those yourself.

Random noise seems to do less harm than chosen noise. Many large language models kept notable performance when nearly all words were shuffled letter by letter, or when many invisible characters were inserted ²¹. Adversarial tests show the other side. On the GSM8K math benchmark, one character edit chosen to do the most damage cut Mistral-7B-Instruct's accuracy from 43.7 percent to 38.6 percent, and eight edits cut it to 19.2 percent ^{2 22}. A dictation error is not chosen to hurt the answer, so it is likely to do less harm than that. A wrong number or a flipped negation in a math or policy question is still worth checking. For the question of whether chatbots cope with such errors, see the question on AI chatbots and dictation errors.

The speech-specific evidence points the same way. A 2024 study found that LLMs were not robust to spoken noise by default, but that training them on a proper set of spoken dialogues could make them more robust ²³. Those models were small and older than today's chatbots, so the finding is about direction, not size.

In practice, the errors that matter most in a spoken prompt are the ones that change a fact the answer depends on. A misheard filler rarely changes the request. A misheard date, product name or instruction can send the model in the wrong direction while the sentence still sounds right. That is why the check after dictation should look for those words first, rather than reading every sentence with equal care.

Listening back is not a reliable check either. In an in-lab study, participants missed identifying over 50 percent of speech recognition errors when they listened to the recognized text ²⁴. Read the transcript before you send the prompt, and check names and numbers first. For how recognition errors are typically measured, see word error rate.

A spoken prompt works best with a goal, context, constraints and the output you want

We suggest a fixed order for a spoken prompt: the goal first, then the context, the constraints and the output you want. The order also means that a small error in one detail leaves the main task intact. For app-specific steps, see how to dictate a prompt.

1. Goal: say what you want in the first sentence, such as "I need to reply to my landlord about the boiler."
2. Context: name the material and the audience, such as the email thread or the person who will read the reply.
3. Constraints: say what must stay the same and what to avoid, such as "keep the dates from her email" or "no threats."
4. Output: state the format and length, such as "a subject line and an email under 150 words."

The same four parts work for a question, a summary or a piece of code. The table shows how each part changes the request.

A spoken version of the landlord request might sound like this: "I need to reply to my landlord about the boiler, which has been off since Monday. Keep the two dates from her email, the twelfth and the nineteenth. Firm but polite, no threats. Give me a subject line and an email under 150 words." Each sentence does one job, so a misheard word in one of them does not break the rest. Numbers and dates are the parts to read back, because a small error there changes the meaning.

Keep the words the request needs and cut the rest. Long spoken prompts often repeat themselves while the speaker works out what they want, and those repetitions add length without adding information.

Code, names and symbols are where voice prompts still break

Spoken prompts struggle with text that does not sound like speech. A 2026 study of spoken questions about code named custom identifiers, such as variable and function names, as a central difficulty for speech input ²⁵. A name such as getUserName can come out as several separate words, and a symbol such as an underscore or a bracket may not appear at all unless you say it.

The same study found a way to recover part of the loss. Refining the recognizer's output with a language model significantly improved performance in both the transcription and the code understanding stages ²⁶. That is the case for a cleanup pass on technical speech. It is still worth reading the refined text yourself, because the study tested its own pipeline and not every chat tool's cleanup.

The fixes are simple. Spell an unusual name out once in the prompt, say symbols by name, and paste exact code, file paths and error messages rather than dictating them. If a name keeps coming out wrong, add it to your tool's dictionary, as explained in how to make dictation spell names correctly. For speaking punctuation and symbols, see dictation commands. For programming by voice, see voice coding. Developers who use coding agents can also read dictation for Claude Code and the voice input guide for coding tools.

Numbers need the same care. A wrong figure changes the answer even when the rest of the prompt is right, so check every figure against your source rather than against the transcript. When a prompt depends on exact wording, such as a negation or a legal term, type that part yourself.

Use the chatbot's microphone for short prompts and system-wide dictation for long ones

For a sentence or two, the microphone button in ChatGPT or Claude is the simplest choice. Our overview of dictation in ChatGPT and Claude covers the buttons, their limits and the known failure modes. If that is enough for you, use it and keep it.

For a long prompt that includes pasted material or several paragraphs of context, system-wide dictation into any text box is easier to manage. You can speak into a notes app or an editor, read the whole draft in one place, and then copy it into the chat. Yap, which we make, is an on-device dictation app for Mac and Windows, and it is one of the tools for this. You can download it to try it, or use the built-in tools if they are enough for your prompts.

Some tools suggest how long to speak at once. Apple's Dictation guide for macOS Catalina 10.15 advised speaking for no more than 40 seconds at a time ²⁷. That guidance is for an older version of macOS, so check the limit in your own version rather than assume it. Speaking in shorter bursts also gives you natural points to check the transcript as you go.

If the request is a question, speak it into the chatbot. If it is a document of context, speak the goal first and the context second. When you want a spoken answer rather than a typed prompt, a voice assistant may be the better tool, because it acts on what you say instead of writing it down. Voice assistants compared with dictation explains the difference. For the wider picture of what AI adds to dictation, see AI dictation, speech to text and LLMs and speech to text without AI.

Voice prompting still struggles with names and identifiers, symbols and code, numbers, and long prompts that are never read back. It works best for short, structured requests with a goal and a clear output, read back once before you send. Speak the prompt, check the parts that are hard to hear, and send the version you have read.