lecturize

I record my lectures and I want them back as notes I can read, not as a two hour audio file. lecturize runs on my machine, transcribes with faster-whisper, picks up where it left off if the run is interrupted, and can translate the result with local models. Nothing leaves the computer except the model downloads.

pipx install lecturize
lecturize run lecture.mp4 --to it

That writes lecture.md, lecture.srt, lecture.it.md and lecture.it.srt next to the recording.

What it does

• transcribes audio and video files (anything ffmpeg reads) with Whisper models, on CPU or on an NVIDIA GPU;
• skips silence with voice activity detection, which is what stops Whisper from inventing text during pauses;
• writes markdown with a chapter every few minutes, plain text, SRT subtitles and Word documents, all from the same timeline;
• resumes an interrupted run. Utterances are saved as they are produced; the next lecturize run on the same file starts from the last one and ends with the same output a clean run would give. lecturize jobs lists what is pending;
• translates into about fifty languages with the OPUS-MT models packaged by the Argos Translate project, run through CTranslate2. Pairs that have no direct model go through English. Translation into English uses Whisper itself, in one pass.

Install

You need Python 3.10 or newer. pipx keeps it out of your other environments:

pipx install lecturize

On an NVIDIA GPU add the CUDA libraries, and the runs get several times faster:

pipx install "lecturize[cuda]"

Models are downloaded on first use into the user cache (lecturize paths shows where): small is 460 MB, large-v3 about 3 GB, a translation package 90 MB. ffmpeg is not required for the common formats, since decoding goes through PyAV, but having it installed does not hurt.

Use

lecturize run talk.mp3                       # small model, language detected, md + srt
lecturize run talk.mp3 -m large-v3 -l it     # a bigger model, Italian spoken
lecturize run talk.mp3 --to en               # English notes, translated by Whisper
lecturize run talk.mp3 --to it -f md,docx    # Italian notes as markdown and Word
lecturize run *.mp4 -o notes/                # many files, outputs in one folder
lecturize run talk.mp3 --fresh               # ignore the checkpoint, start over
lecturize jobs                               # interrupted runs waiting to resume
lecturize languages                          # translation targets available

Options: -m/--model (tiny, base, small, medium, large-v2, large-v3, large-v3-turbo, distil-large-v3, or a folder with a CTranslate2 model), -l/--language, --to, -f/--format, -o/--out, --device auto|cpu|cuda, --compute-type, --no-vad, --fresh.

On my machine (a laptop CPU and an RTX 4070 Ti SUPER) small runs at about three times real time on the CPU in int8 and ten times on the GPU in float16; large-v3 on the GPU stays above real time. A two hour lecture translated into Italian adds about a minute.

The library is usable from Python as well:

from pathlib import Path
from lecturize import device
from lecturize.engine import Engine
from lecturize.pipeline import Job, run

dev = device.pick("auto")
outputs = run(Job(Path("talk.mp3"), model="small", target="it"), Engine("small", dev), dev)

How it is organised

src/lecturize/
  cli.py          the commands (typer)
  pipeline.py     one recording in, files out; resume logic
  engine.py       faster-whisper, streaming utterances with absolute timestamps
  checkpoint.py   SQLite store of utterances per (file fingerprint, settings)
  translate.py    Argos model packages through ctranslate2 + sentencepiece, pivot via English
  text.py         sentences, paragraphs, chapters
  writers.py      txt, md, srt, docx
  device.py       CPU or CUDA, and the DLL paths the [cuda] extra needs on Windows
tests/            unit tests, a resume test with a scripted engine, two slow tests with the real one

pytest runs the fast tests; pytest -m slow downloads the tiny model and runs the real engine on the 37 second synthetic lecture in tests/data.

Licenses

lecturize is MIT. It downloads and runs third party models: Whisper weights converted by SYSTRAN (MIT), and translation packages from Argos Translate, built on OPUS-MT models, which carry a CC-BY 4.0 attribution to their authors. See THIRD_PARTY.md.

MIT © 2026 David Cohen