Lecture Lens: clearer audio, transcripts and summaries for quiet lectures

What I built

Lecture Lens makes a quiet, noisy class recording louder and clearer, and turns it into a Hinglish transcript, a summary, key points and revision questions. It runs entirely on my laptop.

Who I built it for and the problem

My sister. She sits close to the teacher and records every lecture with a voice recorder, because she cannot follow everything in class the first time. She then listens to the recording again at home. She had two problems:

1. The audio was too quiet. The recorder picked up the teacher's voice very faintly, so even at home it was hard to hear.
2. A recording is slow to revise from. To get back to one explanation, she had to listen to the whole lecture again. She needed the lecture as text: a transcript to read, and a short summary with the key points.

So Lecture Lens does both. It makes the audio clearer and louder, and it turns the lecture into a transcript, a summary, key points and revision questions.

Demo

Code

https://github.com/samarth-murade/lecture-lens (MIT license)

How I used open-source AI

• Speech to text: faster-whisper with a Hinglish fine-tuned Whisper model (a community CTranslate2 conversion), so Hindi and English speech come out in English letters.
• Notes: Gemma 4 (through Ollama) writes the summary, key points and 5 revision questions.
• Audio: ffmpeg filters (noise reduction, loudness normalization). This part is not AI.
• Interface: Streamlit.

Why open innovation mattered here

Classroom recordings are private, and a laptop at home may have no reliable internet. Because the models are open, everything runs on my own laptop: nothing is uploaded, there is no subscription, and the models can be swapped for better ones later.

Limits (honest notes)

• It runs on an ordinary laptop CPU with no GPU, so long lectures are slow.
• Transcript quality depends on the recording: good on clear audio, weaker on noisy parts.
• Notes are written by AI and can contain mistakes.

Built with

Python, Streamlit, ffmpeg, faster-whisper, Gemma 4 and Ollama. I built it over a weekend with the help of an AI coding assistant (Claude) and tested it on a real classroom recording.