A one-hour meeting ends, and two weeks later nobody remembers what was decided about the advertising budget. A two-hour tutorial video gathers dust on your "watch later" list. AI does not fully solve this problem, but it can turn a recording into searchable text and a usable summary within minutes.
The main point is that summarizing does not happen directly from audio. First the audio must be converted to text, and then a language model works on that text. Both stages can go wrong, and if the first stage is bad, even a beautiful summary from the second stage will be wrong.
The workflow: from recording to summary#
The whole job can be seen in four steps:
- Record or obtain the file: meeting audio, a podcast file or a video. The cleaner the audio, the easier every later step.
- Transcription: converting speech to text with a speech recognition tool.
- Cleaning the text: correcting names, terms and unclear parts.
- Summarizing: giving the text to a language model with an instruction that defines the output format precisely.
Most people skip step three and then complain about the output. Ten minutes of correcting the transcript usually matters more than switching summarizing tools.
Transcription: the foundation#
Transcription tools come in two kinds: cloud services and open-source models that run on your own computer. The second kind suits sensitive meetings better because the file never leaves your device. We have written more about running models on your own device in running AI on a personal computer.
A few things raise transcription accuracy:
- A close microphone: the distance from mouth to microphone matters more than any software setting.
- One person at a time: overlapping speech almost always transcribes badly.
- A list of special words: if the tool allows, provide company names, product names and technical terms in advance.
- The right language: if the recording mixes two languages, choose the main language correctly and expect more errors on the words from the other language in the middle of sentences.
After transcription, read the text quickly once and fix names, numbers, dates and technical terms. If your tool adds speaker labels ("Speaker 1", "Speaker 2"), replace them with real names; this makes the summary far more useful, because "who agreed to what" depends on those labels.
Long text and the model's memory limit#
The transcript of a long meeting may be larger than the model's context window. If you are not familiar with the concept, see tokens, context windows and hallucination. When the text does not fit in one pass, or fits but the model loses details in the middle, step-by-step summarizing helps:
- Split the text into logical sections, for example ten to twenty minutes each, or one per meeting topic.
- Summarize each section separately.
- Put the section summaries together and ask the model to produce a final summary.
This takes a little time but prevents the middle sections from being dropped. The same logic works for long documents; details are in working with long documents using AI.
A meeting summary prompt#
A vague instruction like "summarize this" gives a generic, useless summary. It is better to define the output format yourself:
The text below is a transcript of a work meeting. Work only from this text and add nothing of your own.
Write the output in these sections:
1. Meeting topic and goal (two sentences)
2. Decisions made (a list; next to each item, one sentence quoted from the text)
3. Action items (each item: what, who, deadline; if the deadline or owner is not stated, write "not mentioned")
4. Open topics and unanswered questions
If a section has nothing in the text, write "not discussed in the meeting".
Meeting text:
[Paste the text here]Two parts of this instruction matter. "Not mentioned" stops the model from guessing, and asking for a quoted sentence lets you check each decision against the original text in seconds.
Video and podcasts: a summary for whom?#
For video and podcasts, the type of summary depends on its use. Decide before writing the instruction:
| Goal | Suggested format |
|---|---|
| Deciding whether to watch | A five-sentence summary and a list of topics |
| Personal note-taking | Key points with approximate timestamps |
| Producing content from it | A list of claims, examples and quotable lines |
| Training a team | Step-by-step instructions and a glossary of terms |
If you plan to make several posts and articles from an interview or webinar recording, a structured summary is a good starting point; the continuation is in idea generation and content calendars with AI. Just remember not to republish other people's content without permission; a summary is for personal use or for reference with attribution, not a replacement for the original work.
Common errors and precautions#
- Hallucination: the model may write a decision nobody made. Direct quotes and "not mentioned" are your best defense.
- Misattribution: when speakers are not separated, the model may attribute one person's words to another.
- Numbers and dates: check every number, amount and deadline against the original recording.
- Tone and sarcasm: jokes and irony get lost in text, and the summary records them as serious.
- Privacy: do not send a work meeting to any service without checking its data terms. We explain what information should not be given to tools in what you should never give to AI.
Summary#
- Summarizing a recording has two steps: audio transcription and text summarizing; the quality of the first step caps the quality of the second.
- Clean audio, a close microphone and quick correction of names and terms have the biggest effect.
- Summarize long text piece by piece, then merge the summaries.
- Specify the output format in the instruction and ask the model for direct quotes for decisions.
- Always check numbers, names and deadlines against the original recording.
- Before sending a work meeting to any service, check attendee consent and the data retention terms.
