How do you turn voice notes into text with AI?

Turning speech into text is a job for a speech recognition model, which listens to the audio and writes out the words. You can run a voice note through one of three routes: the transcript option in your phone's recorder, a transcription app, or an upload of the audio file to a chatbot that accepts audio. Then read the text against the recording before you use it. Summarizing or tidying that text is a separate job for a language model, which our guide to AI voice notes covers next.

Your phone's recorder is the quickest route

On an iPhone 12 or later, Voice Memos can transcribe the speech in a recording. The languages are English in all its variants and nine others, including Spanish, French, German and Japanese ¹. Open the recording and choose View Transcript. Select the text you want, then tap Copy to paste it into a note or a message ².

If the transcript option does not appear, the phone is older than an iPhone 12 or the language is not on the list. The phone must meet both conditions. Our step-by-step guide for iPhone walks through the menus. Samsung's recorder has its own transcript option, covered in how to transcribe a recording on Samsung.

Apps and chatbots can be more accurate, and the result varies

A transcription app or a chatbot runs a speech recognition model too, and the model you get decides the accuracy. In a 2025 test of one 7-minute-31-second podcast episode, Whisper Large V3 Turbo had a word error rate of 1.5 percent, the lowest of the three engines the author tried. Apple's speech API had 10.3 percent on the same recording ^{3 4}.

That is one recording and one author, so read it as a single data point, not a ranking of every app. The word error rate counts the words a model gets wrong, whether it substituted, dropped or added them. A low number on one clip does not promise the same result on your own notes.

Yap, which we make, is an on-device dictation app for Mac and Windows. It turns speech into text as you talk, in the app you are using. It is built for new text, not for transcribing a recording you already made. The question on dictation versus transcription explains the difference.

Accents, conversation and noise change the result

A single speaker reading a script is the easiest case. A voice note is rarely that. A study of conversations between friends found that the best English model, trained on 680 thousand hours of transcribed data, averaged a 19.7 percent word error rate across accent groups. On US English clean read speech, the same model reached 2.7 percent ⁵. Those are different kinds of speech, so the gap does not measure your voice. It does show that a fast, casual, accented ramble needs more checking than a sentence you read aloud.

Background sound adds its own risk. In a 2025 study, researchers played non-speech sounds to Whisper, and a set of hallucinated phrases appeared frequently ⁶. The study used artificial inputs, so it shows the failure mode, not how often it happens in a café. Still, a long silence or a fan's hum can produce words nobody said, so look for any phrase you do not remember saying.

Models also edit what you did not ask them to edit. A 2025 study on dementia speech reports that Whisper tends to omit filler words such as uh and um ⁷. That makes a clean draft easier to read. It is a problem if you need the exact words, such as a quote from your own notes.

Long recordings need splitting

Very long audio is the other common failure. OpenAI says that processing of long audio uploads is best-effort, and that very long recordings may time out ⁸. The upload limit is set in megabytes, not minutes, because file size depends on the format and bit rate. If a recording is very long or near the size limit, split it at natural pauses, transcribe each part, and join the parts in order.

How to get a checked transcript

1. Say the key details clearly. When you record, say each name, date and number once, slowly. This costs a few seconds and saves the most editing later.
2. Make the transcript. Use the phone's transcript option, the app you chose, or the chatbot upload, as in the routes above.
3. Check it against the audio. Play the recording while you read. Fix names and numbers first, because those errors cost the most.
4. Mark what you cannot hear. Write [unclear] where a word is uncertain, rather than guessing.
5. Keep the audio with the text. If you later summarize or make a to-do list, you can return to the recording to check. The question on turning voice notes into a to-do list shows how to do that step.

Other routes work too. Our answers on whether ChatGPT can transcribe audio, using Whisper to transcribe audio and whether Gemini can transcribe audio compare those tools. For the wider subject, see our overview of voice and AI.