How do you turn voice notes into a to-do list?

Make a transcript first, then ask a language model for a task list in which every task names its owner, its due date if one was said, and the words the speaker used. Check each task against those words before you copy it into your task app. The model can miss a commitment, invent one, or assign it to the wrong person, so the list is a draft of your commitments and not a record of them.

Step 1: Get a transcript of the note

A task list needs text, and a speech recognizer makes it. On Samsung's Voice Recorder, select the recording, tap the Transcript assist icon, and tap Transcribe ¹. On a Mac, Voice Memos can recognize and transcribe speech in audio recordings ². Keep the transcript with the recording, because the check in step three needs both.

Step 2: Ask for commitments, not a general summary

A general summary of a note tends to keep the story and drop the commitments. Ask for the job by name, and say what each task must contain:

List every commitment in this transcript as a task. For each task, give the owner, the due date if one was said, and the words the speaker used. If no owner or date was said, write "not stated". Do not add tasks that were not stated.

Each field in that prompt has a job. The owner stops the model from guessing who should act. The quoted words let you find the task in the transcript in a few seconds. The "not stated" rule keeps a missing date from becoming a made-up one.

Step 3: Check every task against the transcript

Read the list with the transcript open. For each task, find the words it cites. If you cannot find them, or if the owner or date is wrong, delete the task or fix it. Then check the reverse as well: read the transcript once more and look for any commitment the list does not have.

The reverse check matters more than it looks, because a summary can drop a commitment that was said only once. In a UC Davis pilot of AI clinical notes made from recorded visits, accidental omissions were the most frequent error, in 18 percent of the 356 reviewed notes ³. Those notes are not voice memos, so treat that number as a reason to read the transcript, not as a rate for your tasks.

Step 4: Use the phone's own suggestions as a starting point

Some phones suggest actions from a recording. Samsung's Voice Recorder with Galaxy AI suggests actions from a transcribed recording, such as adding contacts, events or reminders for tasks ⁴. Treat these as suggestions. Accept only the ones that match the transcript, and add anything the phone missed by hand.

Why the model can get a task wrong

A task list is a rewrite of speech into commitments, and a rewrite can change meaning. A 2026 preprint reports that large language models consistently alter the intended meaning of human writing ⁵. The same preprint finds that even when a model is asked only for grammar edits, it still changes the text in a way that significantly alters its meaning ⁶. That study tested written text, not action items, so it does not measure task lists directly. It does show why a rewrite needs a check.

The common errors are these:

What we could not verify

We did not find a published study that measures how accurately a language model turns voice notes into tasks, with precision and recall we could check. The closest measured evidence is the omission rate above, from clinical notes. If accuracy matters to you, test your own tools: take three recordings whose tasks you know, run the prompt, and count what the list misses and what it invents.

Save the transcript and the list together

Keep the transcript and the checked list in one note, so that the next time you open the task you can see the words it came from. For the summary side of the same workflow, see can ChatGPT summarize a voice memo and are AI summaries of voice recordings accurate. The guide to AI voice notes lays out the whole workflow, and the overview of voice and AI links to the rest of the topic.

A task list is only as good as the transcript under it. The model reads text, not audio, so a word the recognizer heard wrong becomes a wrong task. Read the transcript before you trust the list, and see the large language model entry if you want to know what the model is doing with the text.