Are AI summaries of voice recordings accurate?

Usually, but not reliably. A summary is a rewrite of a transcript, and a rewrite can go wrong in two directions. The model can leave out a point the speaker made, or it can add a point the speaker never made. Published tests find both errors. The studies we could find do not measure summaries of ordinary voice memos directly, so the figures below come from news articles, a public leaderboard and clinical notes. Treat them as a guide to what can go wrong, not as the error rate of your own recording.

Summaries can add content the speaker never said

Summaries can invent as well as omit. In a 2020 study, human annotators reviewed neural summaries of news articles and found "substantial amounts of hallucinated content" in every model-generated summary ¹. Those models are old, so the study shows the failure, not today's rate.

A more current measure is Vectara's public leaderboard. It asks each model to summarize a document using only the facts in that document, and then scores the result. In the September 22, 2026 update, the lowest hallucination rate in the table was 1.8 percent ². The highest was 24.2 percent ³. The model with the highest rate answered only 74.3 percent of its prompts, so read a rate next to the answer rate ³. The test uses one set of documents, so use the table to compare models, not to judge a single summary.

Omissions are the most common error in the clinical evidence

The closest measured picture comes from clinical notes, which are summaries of recorded doctor visits. In a 2024 pilot at UC Davis, physicians reviewed AI-generated notes, and 94.7 percent of the 356 notes were free from significant errors ⁴. The remaining notes had errors, and the most common was omission. An accidental omission appeared in 18 percent of the reviewed notes ⁵, and a hallucination, meaning content the recording did not support, appeared in 11.5 percent ⁶.

A separate study in npj Digital Medicine reported an omission rate of 3.45 percent for LLM-generated clinical notes ⁷. The two studies report omission rates several times apart. The difference suggests that the rate depends on the system, the prompts and how the reviewers counted errors.

Both studies describe clinical notes, not voice memos.

Two things decide how far to trust a summary

Transcript errors carry into the summary. The model reads text, not sound, so a word the recognizer heard wrongly becomes an input the model cannot question. Whisper can also invent text from sounds that are not speech. In a 2025 study, Whisper produced a set of hallucinated phrases frequently when non-speech sounds were played to it ⁸. That study used artificial test sounds, so it shows the failure mode, not how often it happens in daily use. A summary cannot repair a wrong name or number, so check those against the audio.

Length gives the model more to lose. A longer recording has more points to condense, so there are more chances to drop one. This is reasoning, not a measured rate. For a long recording, split it at natural breaks and check each part against its own transcript.

Check a summary against the transcript before you act on it

1. Read the summary next to the transcript, not on its own.
2. Confirm that every name, number and date in the summary appears in the transcript.
3. Search the transcript for decisions and commitments, and confirm each one is in the summary. This is the check for omissions.
4. Delete any line you cannot find in the transcript.
5. For a recording where a mistake would cost money or affect health, listen to the audio at the points that matter, not only the transcript.

The check takes a few minutes, and it targets the two errors the studies above describe. For the wider workflow, read the guide to AI voice notes. The same check applies to the task lists in how to turn voice notes into a to-do list, and to the chat summaries in can ChatGPT summarize a voice memo. If the question is whether your phone can do the summary itself, see can your phone summarize a voice recording. The overview of voice and AI links to the rest of this topic.

The model that writes a summary is a large language model. The recognizer that makes the transcript is an automatic speech recognition system, and its errors reach the summary unchanged.