A few weeks ago, I watched Watson win on Jeopardy! and did a couple of Skype interviews for graduate school. It really got me thinking about machine learning and natural language processing, as I’ve been looking for a way to tie in my web development and programming skills with linguistics in a way that will benefit both fields, as well as “regular” people.
After coming up with and discarding a number of ideas, a shower thinking session led me to a realization: Universal Subtitles is, in effect, building up a corpus of matched video, audio, transcription, and translation! And, because it’s all open, you can remix and play with the data all you want. So, to do just that, I selected the most translated video (which also happened to be the Universal Subtitles introductory video) and began exploring. At first, I looked at a couple of the translations, thinking about doing something with automatic, statistical translation (which is pretty much what Google Translate does, I believe). But I have very little knowledge of that area, so I hit upon another idea: Extracting the text and the audio and matching it up in Praat—automatically.
To do this, I needed the timing information along with the text, and I needed to separate the audio from the video. Doing the latter wound up being the easy part, once I found OggSplit (a part of Ogg Video Tools). I just ran the open Ogg Video file through OggSplit and used Audacity to convert the resulting Ogg Audio file into a WAVE file for Praat to read. But the next part required some work. I remembered from a while back that Hixie and the WHATWG were working on an open subtitling standard, called WebSRT. When I went looking for that, I found that it had been renamed to WebVTT. Universal Subtitles exports its subtitles into many formats, but not yet to WebVTT. Luckily, the renaming of the WebSRT standard to WebVTT was mostly to avoid having to overcome some rather obscure processing issues—the majority of SRT files can become WebVTT files with very little effort. (Namely, adding a “WebVTT” header and converting commas to periods in the timestamps.) I then wrote a script which read the newly-minted WebVTT files and converted them to Praat TextGrids. (This involved “reverse-engineering” the format from TextGrids output by Praat, as I couldn’t find any documentation on the file format itself.)
By the end of this, I had PHP code that could convert unformatted WebVTT files into single-tier TextGrids. I plan to open-source the code, but I can’t decide on a few things: (1) what to call it, (2) where to put it (SourceForge or GitHub), and (3) whether to separate the two out into different projects. (Naturally, the WebVTT parser and the TextGrid generator are in separate modules, but there remains some assumption-based dependencies between them.)
Pairing the WAVE file and the TextGrid revealed some areas for improvement of both Praat and Universal Subtitles. As is inherent in the whole idea of subtitles, dead spaces in the audio (parts without speech) do not get subtitled; thus, the subsequently-generated TextGrid does not fully specify the sections of the tier. Praat doesn’t really like that, as a regular TextGrid specifies even the empty portions of the tier, having positive data from start to finish. Loading a TextGrid that has pieces missing does not make Praat happy, but it still works. (You can practically see the grimace on its face as it sucks it up and makes do.) On the flipside, when matching audio and subtitles up in Praat, you can see how imprecise crowd-sourced subtitling can be. Though it might seem OK when using the web interface to subtitle a video as it plays, when you see the text lined up next to the waveform, it becomes clear that things could be better aligned. (It wouldn’t be that big of a deal, but there doesn’t seem to be any “advanced” interface on Universal Subtitles. There’s not even a place to upload an existing subtitle format, so that you can make corrections locally and upload the new file back to Universal Subtitles.)
Once I had the text and audio lined up, I started extracting formant info from multiple instances of the word “video”. (It appears a total of 11 times in the introductory video, including once in the compound “videomakers” and once in the plural “videos”.) I later found out that about AntConc, and the concept of concordance in general, which helped to identify other words and combinations of words that appear multiple times in the full text of the transcription. (AntConc contains an N-gram viewer, as well.) Long story short, I recorded the formants of the 11 instances of the word “video” and made them available in this Google Docs spreadsheet.
And that’s where the fun started. As you can see in that spreadsheet, I attempted to graph these data points so that I could look at them all together.
Unfortunately, all of the spreadsheet programs I tried (basically everything besides Excel, because I don’t have that) could not graph the data points on top of each other, so that one could fully see the similarities and differences among the formant frequencies. I thought I was going to have to settle for that, but then Michael Newman reminded me that R Can Do Anything™. So I finally had my task to help me learn R.
After a great deal of fiddling around and reading the docs, to not much avail, I headed over see if the folks in #R could help. It turns out they were more than willing to demonstrate how awesome R is; I am particularly indebted to mrflick
, who answered all of my silly questions, one after another. (And, more importantly, and to my amazement, had an answer for all my questions, proving that R really can do anything.) With some more playing around, I was finally able to issue the right commands in the right order and come up with this:
To construct that graph, I still had to somewhat fudge the data by adding an artificial “meter” column so that each instance of the word turned out to be the same width on the graph, but at least R didn’t complain that I was Doing It Wrong™. I also excluded the two instances where the word wasn’t plain ol’ “video”, to avoid creating misleading patterns. I haven’t yet figured out how to automatically calculate the average wave/frequency patterns, but at least now the graph allows you to see them visually. (I suspect I would have to format my data differently to give R enough information to do that properly—right now it doesn’t know that the data is from nine separate instances of the word.)
To top things off, the R script that I used to generate the data is available as a gist on GitHub (released under CC-BY-SA), so feel free to remix the data or improve the script. Just be sure to leave me a comment here telling me what you did! Also, any ideas on where to go from here are welcome!