Skip to content
VarCode-aiPublic

About

A lecture notebook that listens: live speech-to-text with Whisper on OpenVINO, entirely on your computer. A Monad app.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Monad

Lectern

A lecture notebook that listens: live speech-to-text with Whisper, entirely on your computer.

Lectern records from the microphone and transcribes in real time with Whisper on OpenVINO GenAI, on your CPU, Intel GPU or NPU. Notes are organized by course, lecture and category, as plain text files you own.
A Monad app.

Checks Latest release License: GPL-3.0-or-later Runs on Intel CPU, GPU and NPU A Monad app

Monad · Install guide · OVMS Manager · All Monad apps


Lectern recording a lecture: live transcript, waveform and the local GPU indicator

Note

Lectern is part of Monad, a family of desktop apps that run AI models locally with Intel OpenVINO: no cloud, no accounts, no telemetry. If you are new to Monad, start from the install guide: it covers the system packages and the recommended order, OVMS Manager first, then Lectern, so that both apps share the models in ~/models.

Install · Whisper models · Using the notebook · lectern-transcribe · How it works · Privacy · Troubleshooting

Lectern also installs lectern-transcribe, a command-line tool that transcribes every audio or video file in a folder.

Install

Download the AppImage

From the latest release, download Lectern-<version>-x86_64.AppImage, then:

chmod +x Lectern-*-x86_64.AppImage
./Lectern-*-x86_64.AppImage

The same AppImage runs lectern-transcribe, the batch tool:

./Lectern-*-x86_64.AppImage transcribe -i ~/Lectures   # --help for every option

The AppImage carries its own Python, Qt and OpenVINO: it installs nothing, needs no root and runs on any x86-64 distribution with glibc 2.34 or newer. To verify it, compare it with SHA256SUMS from the same release. To add it to the application menu, use Gear Lever or AppImageLauncher, or install from source.

Download the archive and install it

From the same release, download Lectern-<version>.tar.gz, then:

tar -xzf Lectern-*.tar.gz
cd lectern-*/
./install.sh

The installer works as described below. Once the app is installed you can delete the folder.

Install from source

git clone https://github.com/VarCode-ai/lectern.git
cd lectern
./install.sh

The installer copies the files to ~/.local/share/lectern/app and creates a Python environment with OpenVINO GenAI and PySide6 (a few hundred MB). If your NPU driver has no compiler, as in Fedora 44 and 45, it also downloads the official Intel NPU driver into ~/.local/share/lectern/npu-driver (about 50 MB; see the troubleshooting guide). It adds Lectern to the application menu and two commands to ~/.local/bin: lectern and lectern-transcribe. Run ./install.sh --help for every option, and ./install.sh --uninstall to remove it. Your notes and models are kept.

While Lectern records or transcribes a file, the computer does not suspend, not even when you close the lid: a suspend in the middle of a lecture would lose the audio and can hang the GPU. Lectern holds a systemd-inhibit lock only for as long as the work runs. To let the computer suspend as usual, install with ./install.sh --allow-sleep (--keep-awake turns it back on; an update keeps the previous choice). Outside the installer, LECTERN_KEEP_AWAKE=0 does the same, for example for the AppImage.

To try Lectern without installing it, run ./lectern.sh from the clone. It needs a Python environment that already has the packages in requirements.txt.

Requirements: Python 3.10+ and ffmpeg. To use an Intel GPU you also need the Intel compute runtime; see System packages in the Monad README for the package names.

The interface is in English or Italian and follows your system language. To change it, use Tools → Interface language, then restart Lectern. The language also sets the names Lectern gives to new folders and categories (for example Lecture 03 or Lezione 03, Explanation or Spiegazione).

Whisper models

Lectern uses Whisper models already converted to OpenVINO, stored in ~/models (or $OPENVINO_MODELS_DIR). Download them from the Speech-to-Text type in OVMS Manager, or directly from huggingface.co/OpenVINO:

~/.local/share/lectern/venv/bin/hf download OpenVINO/whisper-large-v3-turbo-int8-ov \
    --local-dir ~/models/OpenVINO/whisper-large-v3-turbo-int8-ov

lectern-transcribe downloads whisper-large-v3-turbo-int8-ov by itself the first time if no Whisper model is found.

Model Size Use
whisper-large-v3-turbo-int8-ov ~0.8 GB Live recording. Fast, with accuracy close to large-v3.
whisper-large-v3-int8-ov ~1.6 GB Files. The most accurate on difficult audio, about 3 times slower.
whisper-small-* / whisper-base-* 0.1–0.3 GB Slow CPUs without an Intel GPU. Less accurate.
distil-whisper-* — English only.

Lectern finds every folder under ~/models whose path contains whisper and that holds an openvino_encoder_model.xml. Press F5 to refresh the list after downloading a model.

Which device

Time to transcribe 5 minutes of a real Italian lecture from a file, on a Core Ultra 7 355 (greedy decoding):

Model NPU GPU CPU
whisper-large-v3-turbo-int8-ov 51 s (5.9× real time) 61 s (4.9×) 258 s (1.2×)
whisper-large-v3-int8-ov 109 s (2.7×) 185 s (1.6×) —

Live, on the NPU, a sentence of a few seconds is transcribed in about 0.4 s.

  • NPU (Core Ultra): the fastest on this machine, and it leaves the GPU and CPU free. The first time a model runs on it, it is compiled in a separate process (3 to 6 minutes for the INT8 models; the status bar shows the time) and saved in ~/.cache/lectern/ov_cache. After that it loads in a second or two. Prefer the INT8 models: whisper-large-v3-turbo-int4-ov takes about 24 minutes to compile. On the NPU the course terms are not passed to Whisper and beam search is not used, because OpenVINO GenAI does not support them there yet. If your distribution's NPU driver has no compiler (Fedora 44 and 45), Lectern uses the Intel driver it brings with it: see the troubleshooting guide.
  • GPU (Intel integrated or Arc): fast, with the course terms and High precision. Needs the Intel compute runtime.
  • CPU: works everywhere, about real time with turbo.

Using the notebook

  1. Create a course for each subject. You can save it in the default folder or anywhere else, for example a synced folder. Recurring terms (technical words, proper names) help Whisper spell them correctly in every lecture of that course.
  2. Create a lecture (Ctrl+N). Number and date are suggested automatically.
  3. Record (Ctrl+R). Grey italic text is the preview of what you are saying. At every pause it becomes final text in the normal color and is saved to disk immediately, so a crash loses at most the last sentence. You can correct the final text while the recording goes on.
  4. New transcript (Ctrl+T) splits the lecture by category (Explanation, Exercises, Questions…), even during a recording: from that moment the text goes into the new transcript.
  5. From file… (Ctrl+Shift+R) transcribes audio or video you have already recorded into the open lecture, one transcript per file. This mode is slower than live recording but more accurate.
  6. Export (Ctrl+E) merges the transcripts of a lecture into one Markdown file.

Where notes are stored

~/Documents/Lecture Notes/              # your localized Documents folder
├── .courses.json                       # courses saved in other folders
└── Calculus 1/
    ├── .course.json                    # recurring terms and audio folder of the course
    └── Lecture 03 — 2026-10-01 — Improper integrals/
        ├── 01 — Explanation.txt
        ├── 02 — Exercises — Sheet 4.txt
        └── Recording 10-15-02.wav      # only with "Also save the audio"

Everything is plain files: you can rename, move, back up or sync them with any tool. A lecture folder is recognized by the pattern <word> <number> — <yyyy-mm-dd>[ — <topic>], so folders named in any language (Lecture 02 — …, Lezione 02 — …, Vorlesung 02 — …) show up.

The notebook folder is called Lecture Notes in English and Appunti delle lezioni in Italian. If one of them already exists, Lectern keeps using it whatever the interface language, so switching language never hides your notes.

The recorder dock

The bar at the bottom of the lecture holds everything about recording:

  • Record / Stop and From file…;
  • the elapsed time;
  • a live waveform of the microphone, colored where speech is detected, so you can see at a glance that Lectern hears the speaker;
  • the audio still to transcribe;
  • a Local · GPU chip with the device actually running Whisper (it changes to CPU if Lectern has to fall back): the audio never leaves your computer;
  • the status, and the recording settings.

Appearance and reading

  • Tools → Appearance: theme (System, Light or Dark: System follows the desktop, also when it switches at sunset) and Reduce motion, which keeps the status dot steady.
  • The Aa button above the transcript sets the reading size (small, medium, large) and the typeface: Sans serif (Inter) or Serif (Newsreader, comfortable for long reading).
  • Ctrl+K opens the command palette: type a few letters to run any menu action, open any lecture of any course, or start a new transcript in a category.

Recording settings

Open them with Settings (Ctrl+,).

Setting Meaning
Microphone Input device. The list updates when you plug devices in.
Model Whisper model. Use a turbo model live.
Device GPU, CPU or NPU. The first time a model runs on the NPU, compiling it takes a few minutes; later loads take about a second. If the device cannot be used, Lectern switches to the CPU and tells you. If the GPU fails twice during a lecture, it continues on the CPU without losing audio, and a banner offers to restart Lectern on the GPU: the recording starts again by itself.
Language Language of the lecture. Auto-detect works, but a fixed language is more reliable on short sentences.
Precision Fast (greedy decoding) is recommended live. High (beam search 5) is more accurate but every sentence arrives a few seconds later.
Sensitivity, Pause How loud speech must be, and how long a silence closes a sentence. For a distant speaker use High sensitivity and a pause around 1 s.
Context Terms valid for every course, added to the course's recurring terms. Not used on the NPU, where OpenVINO GenAI does not support them yet.
Preview while you speak Turn it off to save power on battery: text appears only at each pause.
Time for every sentence Prefixes each sentence with [hh:mm:ss].
Also save the audio Saves a .wav in the lecture folder, or in the course's audio folder if set. You can transcribe it again later with a more accurate model.

Keyboard shortcuts

Keys Action
Ctrl+K Command palette: actions, lectures, new transcripts
Ctrl+N New lecture
Ctrl+T New transcript
Ctrl+R Start / stop recording
Ctrl+Shift+R Transcribe audio files
Ctrl+E Export the lecture to Markdown
Ctrl+Shift+C Copy the transcript text
Ctrl+O Open the folder
Ctrl+, Recording settings
F5 Refresh the model list
F1 Quick guide

When a new version is released, a banner at the top of the page links to its notes. Help → Check for updates checks right away, and Help → Check for updates automatically turns the daily check off.

Batch transcription: lectern-transcribe

lectern-transcribe transcribes every audio or video file in a folder (mp3, wav, m4a, aac, flac, ogg, opus, wma, webm, mp4, mkv, mov) into .txt files in a transcripts subfolder. Files already transcribed are skipped, so an interrupted run resumes where it stopped.

lectern-transcribe                       # interactive: asks for folder, device, model, language…
lectern-transcribe ~/Lectures GPU        # no questions, default settings
lectern-transcribe ~/Lectures GPU -l it -c strong -t --prompt "Calculus: derivatives, integrals"

From the third argument on, the options are passed to transcribe.py, which you can also run directly:

~/.local/share/lectern/venv/bin/python transcribe.py --input ~/Lectures --device CPU --fast
Option Default Meaning
-i, --input DIR current folder Folder with the files.
-r, --recursive off Include subfolders. The output mirrors the folder structure.
-o, --output DIR <input>/transcripts Where to write the .txt files.
-d, --device GPU GPU, CPU or NPU. Falls back to the CPU if the device cannot be used.
-m, --model large-v3, then turbo A local model folder or a Hugging Face repo id. Missing repos are downloaded to ~/models.
-l, --language auto ISO code (en, it, de…), language name, or auto.
-c, --cleanup light Audio filters: none, light (removes hum and hiss, levels the voice), strong (also reduces background noise).
-b, --beam N 5 Beam search width.
--fast off Greedy decoding (--beam 1), 3–5 times faster.
--prompt TEXT none Subject, terms and names. If the model starts repeating itself, the file is transcribed again without the prompt.
-t, --timestamps off One line per sentence, prefixed with [hh:mm:ss]. Without this option, text is split into paragraphs at long pauses.

How live transcription works

  1. The microphone is recorded as 16 kHz mono through PipeWire's pw-record (package pipewire-utils or pipewire-bin) and goes into a buffer. Without pw-record, or with LECTERN_AUDIO=qt, Lectern records through Qt instead; on some laptops Qt 6.12 gets audio in which hardly any speech is detected. A watchdog reopens the microphone if it stops sending audio, which can happen with PulseAudio/PipeWire after a device suspend.
  2. Silero VAD, run with OpenVINO on the CPU, gives the probability of speech for every 32 ms frame. If Silero cannot be downloaded on the first start, a simpler volume threshold is used instead.
  3. A worker thread cuts the audio into sentences at pauses. Segments are kept between 1.5 and 10 seconds long, because Whisper skips words on longer ones.
  4. While you speak, the current segment is transcribed greedily for the preview, with a time budget so it never delays the final text. At each pause the segment is transcribed for good and appended to the file.
  5. Whisper hallucinations are filtered out: loops (the same phrase repeated), phrases it invents on silence such as "Thanks for watching", output that just repeats the prompt, and output suspiciously short for the amount of speech.

Files (From file… and lectern-transcribe) use Whisper long-form transcription instead, with timestamps and ffmpeg audio cleanup. Each 30-second window keeps the previous text as context, which is more accurate than transcribing short sentences one by one.

Files

Path Content
~/Documents/Lecture Notes/ Your notes (see above).
~/.config/lectern/lectern.conf Settings and the last open lecture.
~/.cache/lectern/lectern.log Diagnostic log: timings and decisions of each step (Help → Open log file).
~/.cache/lectern/silero_vad_openvino_16k.onnx Voice detection model, downloaded on first use.
~/.cache/lectern/ov_cache/ Compiled GPU kernels, which make later starts faster.
~/.local/share/lectern/ Installed app and Python environment.

Tips and troubleshooting

Problem Fix
"to transcribe" keeps growing during a lecture The model is too slow for this device. Use a turbo model, Fast precision and the GPU, or turn off the preview.
Words are lost when the speaker is far away High sensitivity, pause around 1 s, and add the course terms. Turn on Also save the audio and transcribe the .wav again later with large-v3.
The text repeats the course terms instead of the speech Lectern detects this and retries without the context. If it keeps happening, shorten the recurring terms.
Speech is almost never detected (the waveform moves but is rarely colored) Install pw-record (pipewire-utils on Fedora, pipewire-bin on Debian/Ubuntu). Help → Open log file shows which way the microphone is recorded (via pw-record or via Qt).
No microphone found Check that PipeWire or PulseAudio is running and that the microphone is not muted in your system settings.
The GPU is not used Install the Intel compute runtime (see System packages) and check Help → Open log file.
The NPU fails with ZE_RESULT_ERROR_UNSUPPORTED_FEATURE The NPU driver of your system has no compiler, as in Fedora 44 and 45. Lectern then uses the Intel driver it brings with it: with the installer, run ./install.sh again so that it downloads it. See The NPU is listed but the model does not load.

More fixes are in docs/troubleshooting.md, and the problems common to every Monad app in the Monad troubleshooting guide.

Privacy

Lectern transcribes on your computer: your recordings, transcripts and notes never leave it. It connects to the internet only to download the Whisper model you choose, the Silero voice detection model on the first recording (1.3 MB; offline it falls back to a volume-based detection), and to check at most once a day for a new release, which you can turn off in Help → Check for updates automatically. The full list of connections, for every Monad app, is in Private by design.

Contributing

Bug reports, translations and pull requests are welcome: see CONTRIBUTING.md.

License

Copyright © 2026 VarCode-ai. Lectern is free software, released under the GNU General Public License v3.0 or later. Third-party components keep their own license: OpenVINO GenAI (Apache 2.0), Qt for Python / PySide6 (LGPL v3), Silero VAD (MIT), the Whisper models (MIT), and the Inter and Newsreader fonts (SIL Open Font License 1.1, see common/fonts/).


Monad
Lectern is a Monad app · local AI for OpenVINO

About

A lecture notebook that listens: live speech-to-text with Whisper on OpenVINO, entirely on your computer. A Monad app.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages