From a 75-minute lecture recording to study-ready notes

Lecture transcription turns a 60-90 minute classroom recording into searchable text plus a topic-grouped summary in about 12 minutes total — roughly 10 minutes for the transcript and 2 minutes for the AI summary pass. Accuracy lands around 92% on a clearly-miked professor speaking a high-resource language, and drops on heavy accents or dense technical vocabulary unless you correct the terms once and reuse them. Check your institution's recording policy before uploading — lectures are often the professor's or university's property.

The workflow, end to end

You finish class, stop the recording on your phone, and have a 75-minute M4A or MP3 file. Drop it into the browser uploader or paste the URL if the lecture is already on YouTube or a course portal that yt-dlp can reach. No app install. The file uploads, transcription starts, and you can close the tab — the job runs server-side.

About 11 minutes later you have a transcript and an AI summary. The transcript is a single-speaker bubble layout (the professor) with timestamps every paragraph. The summary is 6-10 bullet points grouped by topic, generated from the full transcript by a separate LLM pass. Click any bullet to jump to the matching moment in the transcript. Export to DOCX for annotation, TXT to paste into your notes app, or JSON if you want to parse it.

The whole loop fits inside the time it takes to walk home and make coffee.

How accurate is lecture transcription, really?

For a professor speaking clearly into a lapel mic or directly into a phone placed on the desk, expect around 92% word accuracy — the same plateau that holds for podcast audio at 128 kbps or higher. That's roughly one error every 12-15 words, mostly on proper nouns, acronyms, and discipline-specific terms.

Three things drag accuracy down on lecture recordings specifically:

Distance from the speaker. A phone in your backpack 8 metres from the lectern can lose 10-15 points of accuracy compared to the same phone on the front row.
Strong non-native accents in the lecture language. The model still handles them, but expect more substitutions on technical terms. The conversational filler stays accurate.
Domain vocabulary. A microbiology lecture full of Pseudomonas aeruginosa and quorum sensing will get those terms wrong on first pass. So will an EE lecture on MOSFET topologies or a philosophy seminar on Heideggerian hermeneutics.

The vocabulary problem has a workable fix. After the first lecture of a course, scan the transcript for the 10-20 terms it consistently misheard, and keep that list. For subsequent uploads, do a find-and-replace pass in the exported DOCX — or, if you're using the REST API, pass a custom vocabulary list with the request. The model gets the terms right from the start of the second lecture onward.

Try it on your audio

Start free →

30 minutes a month, no card.

What about accented professors?

Accuracy on accented speech is the question students ask most, and the honest answer is it depends on the accent, not on whether there is one. A professor who learned English in Mumbai or Lagos and lectures fluently will transcribe at roughly the same accuracy as a native speaker. A professor whose L1 phonology is very distant from English (some Slavic, East Asian, and Arabic L1 backgrounds) and who speaks at speed may drop to 85-88% — with the errors concentrated, again, on technical vocabulary rather than connectives.

If your professor lectures in their L1 — German, French, Mandarin, Korean, Arabic, Russian, Spanish, Portuguese, and 90+ others — let auto-detect run on the first 30 seconds. One price across all 99 languages, no tier system. The transcript ends up in the lecture's actual language, which is what you want.

The permission question

Before uploading: most universities treat lecture recordings as the professor's intellectual property, the institution's, or both. Some courses explicitly prohibit recording; some allow it only for personal study; a few publish lectures openly. Before you upload, check:

The course syllabus or the institution's academic policy on recording.
Whether your registered disability accommodations cover recording (in many jurisdictions they do, with conditions).
Whether the professor said anything at the start of term — many state their policy in week one.

If the answer is "personal study only", that's compatible with using lecture transcription as long as you don't redistribute the transcript. Privacy-wise on our end: source audio is permanently deleted from infrastructure within 24 hours of job completion. Transcripts stay in your account until you delete them. We do not train models on your data. So the audio doesn't linger after the run — but the upload itself is still your call to make against your institution's rules.

If your course explicitly prohibits recording, don't.

What this means for your study workflow

For most students taking a lecture-heavy course, the practical setup is:

Record the lecture on your phone (Voice Memos on iOS produces M4A at AAC; the default Android recorder produces M4A or MP4). Place the phone within 2-3 metres of the speaker if you can.
After class, upload directly from the file. A 75-minute lecture is around 35 MB at standard quality — well inside the 100 MB free-tier cap and trivially inside the 2 GB Pro cap.
Read the AI summary first, then skim the transcript for the bullets that matter. Most students don't need to read 75 minutes of text — they need the 8 topic chunks and the ability to jump to any of them.
Keep your course vocabulary list. After lecture 1 of a course, fix the 15 terms the model got wrong. Lecture 2 onward gets them right.

On the free tier you get 30 minutes per month — enough to evaluate the result on one short recording, but not enough for a full course. Pro at $19/month covers 600 minutes, which is about 8 lectures at 75 minutes each, with overage at $0.04/minute on Pro and $0.02/minute on Business beyond that. See pricing for the comparison.

FAQ

Can I transcribe a lecture from a course portal video link?

Often yes — if the video is on YouTube, Vimeo, or one of the 1,500+ platforms supported by yt-dlp, paste the URL into the uploader and we'll fetch the audio. If the lecture is on a closed course portal (Canvas, Moodle, Blackboard) that requires login, the URL won't work — download the video file from the portal first, then upload the file directly. Check your institution's terms before doing either.

How long does a 75-minute lecture take to transcribe?

Approximately 6× faster than realtime — a 75-minute lecture completes in about 12-14 minutes including the AI summary pass. The transcript itself is usually ready in 10-11 minutes; the summary adds another 1-2 minutes. You can leave the browser tab and come back; the job runs server-side and email-notifies on completion if you set that up.

Will the transcript label the professor vs students who ask questions?

Yes, on mono recordings the speaker diarization model labels distinct voices as Speaker 1, Speaker 2, and so on. You rename them in the UI by clicking the speaker chip — typically Speaker 1 becomes "Prof. Smith" and student questioners stay as Speaker 2, 3, etc. On a stereo recording with the professor on one channel and the room mic on the other, the split is automatic and exact.

Does the AI summary actually understand lecture content, or just paraphrase?

It groups the transcript into topic chunks and pulls the key claim or definition from each chunk, plus any explicit "remember this for the exam" cues the professor said out loud. It does not fact-check, fill in missing context from outside the lecture, or solve worked problems. Treat it as a study aid that surfaces what was said, not as a substitute for thinking through the material.

What file format should I record in?

Whatever your phone records natively is fine. Voice Memos (iOS) produces M4A at 64 kbps AAC, which transcribes accurately. The default Android recorder produces M4A or MP4. For a longer or more important lecture, switch to a recording app that writes WAV or MP3 at 128 kbps or higher — the accuracy plateau sits at 128 kbps, so going beyond it does not help, but going below 64 kbps starts to cost a few percentage points.

What if the professor uses a lot of equations or technical terms?

First lecture: expect the model to mishear domain terms. After lecture 1, build a 10-20 word vocabulary list of what it got wrong (substitutions are usually consistent — MOSFET might come out moss fit every time). For subsequent lectures, either do a one-pass find-and-replace in the exported DOCX, or pass the vocabulary list to the REST API on submission. From lecture 2 onward, the terms come back correct.

Is my lecture recording private after I upload?

Source audio is permanently deleted from infrastructure within 24 hours of job completion. Transcripts stay in your account until you delete them. We do not train models on your data. Separately from our handling: your institution's policy on recording lectures still applies — privacy on our end doesn't override a course rule that prohibits recording in the first place. Check the syllabus.

Lecture transcription: 75-minute recording to study notes