Pop Up Archive: Breakthrough in Speech to Text?
For the past year I've been working with a startup called the Pop Up Archive to improve the way WILL is archiving digital audio. It's one thing to save a number of files in some form of managed digital storage - and in WILL's case it's a very large number of audio files. The challenge is knowing what the files are. Metadata is the yin to a bitstream's yang; without it, we can't really claim to be preserving anything.
Why should we care about preservation? Because media technologies and formats are changing so fast, the only way to keep things accessible is to manage them over time.
So we have many thousands of digital audio files and we need metadata. We typically create a basic title for each asset, and keep things like creation date, producer names, and usually a short description and some subject terms. But if you were looking for anything more granular, or if you were looking for Google to be able to search for anything significant contained in the audio itself, you would be disappointed.
This is where the Pop Up Archive comes in. They are developing next generation speech-to-text technology that appears to represent a breakthrough in accuracy. Automatic transcription of audio is one of those hard problems. The audio quality can vary from great to terrible. Music or noisy background environments can obscure the words. Accents and specialized vocabularies can complicate things further.
But check out this transcript for one of WILL's audio archives:
https://www.popuparchive.com/collections/1472/items/10811
Here's one of a recent WILL News story:
https://www.popuparchive.com/collections/1468/items/29493
Here's an interview with the writer Jean Thompson talking about her recent book The Witch and Other Tales Re-told:
https://www.popuparchive.com/collections/1468/items/28294
Are these perfect? Not yet, but the improvement over anything I've seen before is astonishing. Most importantly, they are good enough to make the audio content indexable by Google, and accessible to people with impaired hearing. The full meaning of the audio content is accessible.
Let's just use the word "accessibility" to mean universally usable by all. That's what we're trying to do with public media. If we're going to do that with audio, we need to extract the semantic content from the bitstream and make it accessible. A year ago, we signed up with the Pop Up Archive to try and make that happen. They're moving forward faster than I imagined.