speech_enhancement

This project is a set of Google Colab notebooks for cleaning up noisy audio in a video, then putting the cleaned audio back onto the original footage. Colab is Google's free service for running code notebooks in the browser with access to a GPU, so nothing needs to be installed on the user's own computer.

There are two notebooks with different jobs. The first is an experiment notebook that runs six different audio cleanup pipelines on the same source video and produces a comparison for each one, mixing tools like DeepFilterNet, Resemble Enhance, LavaSR, and FFmpeg's loudness normalization in different orders. Each successful pipeline produces its own comparison audio file, and the ones that also rebuild a video produce a matching video file too, so a person can listen to each version side by side and judge which sounds best. The second notebook is a simpler, focused version that skips the comparison step and just runs the one combination the author settled on, DeepFilterNet followed by Resemble Enhance, then writes a new video file that keeps the original picture but swaps in the enhanced audio track.

To use either notebook, a person opens it in Google Colab, switches to a GPU runtime, runs the setup cells to install the needed libraries, then uploads a video file or points the notebook at a file already stored in Google Drive. After the cleanup runs, the notebook shows a table for comparing the results and lets the results be played back directly in the browser before downloading them as a zip file.

The repository also includes a couple of small test scripts that check the notebooks are still structured correctly, plus documentation describing the design notes and plan used while building the project. Any downloaded videos or generated audio files are deliberately excluded from the repository itself, so only the code and notebooks are kept in version control.