Took me a minute to get that the first one was a keyword, and the second one actually got me an audio clip of someone saying something.

If the problem you are trying to solve is easing someone finding someone saying a word, then I fell it might be easier for the user to just have someone use the "second box" (transcript search) to do all of the searching, and then narrow based on either suggested keywords or some other offering.

Also I noticed a fairly minimal amount of videos when doing pretty big searches (car), any reason to limit it to 5ish videos?

It would also be interesting if you could chain together sentences somehow.

I like it!

You should do a regular youtube search concurrently and remove its results from yours.