This post includes more technical details on a longer post I have included on the Sounding Out blog in which I mention that we analyzed the recordings in the UT Folklore Center Archives at the Dolph Briscoe Center for American History, The University of Texas at Austin, which comprises 57 feet of tapes (reels and audiocassettes) and covers 219 hours of field recordings (483 audio files) collected by John and Alan Lomax, Américo Paredes, and Owen Wilson, among others. We wanted to find different sonic patterns including the presence of instrumental music versus singing versus speech. The results of our analysis are noteworthy. For example, in the visualization shown in this brief movie, we see a subtle yet striking difference between the Lomax recordings (created 1926-1941), which are the oldest in the collection, and the others created up until 1968. The Lomax recordings (primarily created by John Lomax) consistently contain the least amount of speech in comparison to the other files.
UT Folklore Collection, Visualizing the predicted presence of Instruments, Speech, and Song using ARLO from Tanya Clement on Vimeo.
How was this data produced? We used the ARLO software. We tagged 4,000 randomly selected two-second windows; ARLO divided these windows into 1/32 windows.
We ended up with 93966 instrument tags, 48718 spoken tag and 81890 sung tags. With all the spectra tagged (even non-instrumental, speech, or sung), we had 25,053,489 (all spectra, all 4,000 files).
The results in the movie are shown for each file, grouped according to date across the x-axis. The dates are shown at the top of the screen. The Y-axis shows the number of seconds that each class (green=instrumental; red=spoken; and purple=sung) was predicted highest for each file. The blue bar shows the total number of seconds for each file. The movie shows a scrolling of these results across the collection according to date.
Of course, there are a number of ways you can read these results, which I’ve outlined on the longer post on the Sounding Out Blog.