Week 6 – Audio Alteration Success
This week, I played around with the noise reduction library I found last week and began using it for actual audio from old movies. I also revisited my colorization model to see if I could improve it further.
Noise Reduction Results
The noise reduction library came with two methods of noise reduction: stationary and non-stationary. Stationary noise reduction eliminates background static that is focused around a similar mean frequency (the static generally stays the same). Non-stationary noise reduction eliminates all kinds of static from different frequencies (usually different forms of background noise).
For my project, I will be using stationary noise reduction. To test it out, I used the audio from an old Indian movie. Unfortunately, I cannot upload it to WordPress because of conflicting file types. So, to give you an overview of how the noise reduction method worked on this sound, the background static that was present throughout the audio was eliminated while preserving music and character voices in the scene – a success!
This week, I tried running my colorizer model on more training samples. The resulting colorized video was significantly worse than what I had been using so far. Color contrasts between objects were so extreme, the video still resembled a black-and-white video for the most part. With no more ideas on improving the colorization, I will have to close the door on this stage of the project and simply move on.
With audio alteration and graphical alteration done for the most part, the final stage is simply to merge the two and begin working on my final product. The first step of this stage is to find a way to split input movies into video files and audio files for each model to work with.