I have a knack for remembering audio. I’m awful at remembering names and faces, but if I hear something I can often recall it later. This has manifested itself as a bit of a party trick for the podcasts I listen to, where I can quickly find the section of a show where a topic was discussed even years after I heard it. Fun, but not particularly useful.
This situation gave me the idea for a little side project, PodSearch, empowering the same quick podcast recall for anyone. The concept was simple. Take a few of my favorite podcasts and run them through automated speech-to-text and make the result searchable.
At first I wasn’t sure what to expect from fully automated speech recognition. The results are pretty comical:
The Incomparable, Number 324, January 2017. Welcome back everybody to The Incomparable, I’m your host Jason Snell. So Batman won a tournament that we ran here that was stupid, and we honor Batman by doing episodes about Batman.”
the incontrovertible numbers we January 2n474 welcome back everybody to be uncomfortable on your host Jason L we so bad man won the tournament that we ran here that was stupid and we wanted to have an amazing episode about that man
This is absolute gibberish.
But after playing with the output a bit I found that even though this is an awful transcription, it is actually pretty good for keyword searching. While the context is usually lost and words are sometimes wrong (“Batman” -> “that man”), it gets it correct enough to be useful.
I ran the resulting scripts through a handful of podcasts and have published the result on PodSearch.
You can easily search for a term or keyword and then play the actual audio back to find if it was the section you were thinking about. I even tag the sections with timecoded Overcast links for easy sharing. It is very rough and a work in progress but surprisingly fun to play around with.