Home → Blog → How to turn a lecture or webinar into usable notes

How to turn a lecture or webinar into usable notes

Published 2026-08-20 · 5 min read

A ninety-minute lecture is roughly 12,000 words. Rewatching it for one formula is a waste, and taking notes by hand costs as much time as the recording itself. A transcript solves this in minutes: you can search it, condense it into notes and come back to it before an exam or a rollout.

Why the text version wins

A transcript covers the three reasons people rewatch recordings in the first place:

Search. A term, a library name or an author is one keystroke away instead of a scrubbing session.
Notes. Condensing written text into a summary is a single pass, by hand or with a language model.
Accessibility. People read when listening is impractical: on a commute, in an open office, with hearing loss.

Course authors get a bonus: a page with a transcript is indexed by search engines, while a video is not.

Prepare the recording

Lectures almost always arrive as video. Strip the audio track and compress it — recognition does not need the picture, and the extra megabytes only slow the upload:

ffmpeg -i lecture.mp4 -vn -ac 1 -ar 16000 -c:a libopus -b:a 24k lecture.ogg

Mono, 16 kHz and opus shrink the file by an order of magnitude with no loss of intelligibility: the model resamples to 16 kHz anyway.

Transcribe with timestamps

For a lecture the segment format pays off — it lets you jump back to the exact minute of the video.

from openai import OpenAI

client = OpenAI(base_url="https://voicesscribe.com/v1", api_key="your key")

with open("lecture.ogg", "rb") as f:
    r = client.audio.transcriptions.create(
        model="whisper-1", file=f, response_format="verbose_json",
    )

for seg in r.segments:
    print(f"[{int(seg['start'])//60:02d}:{int(seg['start'])%60:02d}] {seg['text'].strip()}")

You get text split into utterances with time marks. Publish it next to the video and every timestamp becomes a link into the recording.

Terms, names and formulas

Every lecture is full of domain vocabulary, and that is exactly where any model slips. Pass the list as a prompt:

r = client.audio.transcriptions.create(
    model="whisper-1", file=f,
    prompt="Econometrics lecture: heteroskedasticity, Durbin-Watson statistic, OLS",
)

The prompt never shows up in the output — it only nudges recognition toward the right words. If the speaker sticks to one language, pass the language code as well: it removes detection errors on pauses and filler phrases.

Handling the full ninety minutes

A long recording will not fit into a single request, so cut it at pauses and stitch the results:

ffmpeg -i lecture.ogg -f segment -segment_time 900 -c copy part%03d.ogg

Fifteen-minute chunks process in parallel and merge back in file-name order. Prefer cutting on silence so no sentence is torn in half.

Try it on your own recordings. Sign-up takes a minute, and the free minutes are enough to judge the quality.

Get a free API key

Frequently asked questions

How long does an hour-long lecture take?

Processing runs faster than real time: an hour of audio comes back in a couple of minutes, and splitting it into parallel chunks makes it faster still.

Does it work with poor lecture-hall audio?

Speech close to a microphone transcribes reliably. A reverberant hall, questions from the back rows and projector noise cost accuracy — a lapel mic or a feed from the desk fixes most of it.

Can I get subtitles for the video directly?

Yes, request srt or vtt and the response is a ready subtitle file that plugs straight into a player.

How do I separate the lecturer from audience questions?

There is no automatic speaker separation. If questions go through a separate microphone recorded to its own track, transcribe the tracks separately.