I'm working on studying an endangered language and have the following problem: while I already have labeled exactly what the speakers I worked with have said at given timesteps on the word level and I know what possible lists of sounds/phones/letters can make up the word, I need a way to automatically detect the boundaries of the sounds with minimal data (maybe a few hours of recording).

To handle the case where a word may only be pronounced one way at first, I wonder if finding the difference or angle between MFCC vectors to create a function of MFCC change and then finding the max n points could work, but I'm not sure if there is a better way of doing this. I also worry about the fact that some boundaries may be less pronounced between sounds than others. I also don't know how to handle the case where a word may be pronounced with varying numbers of phones (eg. "texts" /t eh k s t s/ -> [t eh k s] or [t eh k s t s]).

Any help would be appreciated

  • 1
    $\begingroup$ This sounds like a pretty cool project! $\endgroup$
    – mmmm
    May 25 '21 at 23:02
  • $\begingroup$ what'sa "MFCC"? $\endgroup$ May 26 '21 at 4:04
  • $\begingroup$ i s'pose it's this. Google (and Wikipedia) are often my friends. $\endgroup$ May 26 '21 at 4:06
  • 1
    $\begingroup$ If you have raw audio data, consider scattering. It's the SOTA on limited data and goes steps further than MFCC. $\endgroup$ May 26 '21 at 4:33
  • $\begingroup$ Awesome, I'll check it out! Do you know of any other similar metrics I could use @OverLordGoldDragon? Even if it may be viewed as "worse" or better in specific contexts, comparisons would be useful $\endgroup$
    – cmitch
    May 27 '21 at 22:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.