The Engineered Network: Transcriptions

2 November, 2018

(Originally published 20th June, 2018) As real-time processing power improves and server-side software can search audio for specific words or phrases on demand or build a catalogue in the background for text search of audio files en-masse, like google does now for text and to some extent also for photos, transcriptions of podcast audio will become obsolete. That said, such a reality isn’t within our grasp just yet, with the best audio transcription software not much better than Siri, Alexa or Google Home assistants. Inspired by the ever-creative “underscore” David Smith and his podcast search side project (his words), and seeing the utility in text-driven audio search, I decided it was time to do something similar at TEN.

Like David, I’m happy with approximation. Of course I could spend countless hours manually transcribing hundreds upon hundreds of hours of audio, or I could rely on transcription software to handle this for me and give its best effort. It isn’t perfect, but it’s good enough to pick out keywords and phrases to allow you to hone in on roughly when and in what episodes certain topics were discussed. Unlike David, I didn’t use a series of scripts into a web service backend to do the heavy lifting, instead I relied on a commercial off the shelf tool that I’d forgotten I still have a copy of from a previous life: Dragon Dictate. The interesting thing I found about the software was that the older Windows version (v13) was the only one that was stable and compatible enough with all of the MP3/AAC files and formats I’ve used over the past 5 years. Attempts to use the Mac version of Dragon failed miserably, with it crashing constantly and refusing to transcribe episodes for reasons unexplained. To accomplish this Dragon v13 was run up in a Windows 7 VM and set to auto-transcribe. The entire backcatalogue of TEN shows took nearly a week of constant background processing to complete, but the results are in.

As of today the rough-transcripts can be downloaded as text files if you like, from a link at the bottom of each podcast episode page, and under the Search option for the website it should pick up which episode mentions which words via DuckDuckGo. I’ve only rolled out Pragmatic episodes for the moment but the others will follow in coming weeks. As for motivation, I’ve had several listeners ask for this feature for quite some time, but to be honest this solution is just as much about me. Having recorded and published hundreds of hours of audio it’s become clear to me that my memory is sometimes at variance with the facts of my previous recordings and when I’m covering a topic I want to firstly make sure I haven’t covered it previously, and if I have, to be sure I’m not repeating myself or I can at least refer to it correctly.

Hopefully it’s a useful feature. Enjoy!