We want busy people to save time sifting through videos and watching redundant content so we created Video Bento!
How Customers Use Video Bento
Our users are looking for videos by doing a search: 1) Users enter a Youtube search term 2) Video Bento provides - A list of enhanced topics - A filtered list of videos that succinctly captures the essence of the query
How we built it
Several components make up Video Bento:
- FE/BE with React and Flask
- Various python tools that work with content from Youtube
- Finding Search Results
- Obtaining video transcripts
- Pytorch Elements to find similarities between video topics and content
- Summarizer - An RNN-based encoder-decoder for abstractive summarization
- Topic Encoding and Distance measurements
Challenges we ran into
- The data pipeline: getting raw video to tokenized data for input into our models. Particularly in obtaining accurate transcripts of youtube videos. We attempted to use DeepSpeech for this, but could not get it to work with any significant accuracy. Finally, we decided to use a library: youtube-transcript-api.
Accomplishments that we're proud of
- We have a deployed tool with a nice UI !!!
What we learned
- We faced the challenges and nuances of ranking and sorting videos with respect to content and topic.
What's next for Video Bento
- The next step is to provide a summarized video, which pulls from the most relevant snippets of several videos of a similar topic and stitches them into one single video.