As an experiment and exploration, it was very useful to let the program run, shoot out a block of text, and generate the resulting timing blocks for the purposes of subtitles or transcription:
…but it would have been foolhardy to 100% walk away and let it do transcription without a second human-driven scan through the results to find mistakes. I’ve been that human, and I’ve seen things.
I’ve seen the resulting transcriptions do great jobs with proper name capitalization, odd and challenging punctuation, and paragraph breaks. I’ve also seen it knock itself silly on my New York accent and non-obscure phrasing, and definitely making a poor guess on my made-up word “Cowicature”. The algorithm works great, except when it doesn’t.
And here we get to a turn of phrase I’ve come to adapt, which is an alternate term for AI: “Algorithmic Intensity”. The human need to give life and will to machinery is a very long-lived one; but most who phone number database look at the code behind this mechanism would agree – it’s just code. The only difference is that the amount of computing power and data to derive the outcome dwarfs numbers considered unattainable a decade or two ago.
And the speed can’t be beat – 5 years of weekly podcasts took roughly 4 hours to completely transcribe by Whisper, and that amount of time was simply because it was set as a lower-priority action in the queueing system. time for all the uploaded episodes of my podcast are sitting in that collection is a little involved, but my back of napkin estimation is it’s two and a half solid days of me talking about technology and emotion, ranging subjects from programming and compilers to summer camp and family. I can assure you – I was not going to transcribe these podcasts anytime soon, and I was not going to ask someone to do it. While I’m proud of my work, I’m in no position to be able to record four podcasts in a month and create paragraphs of text from them.