I've been collecting speech data for my baby brother (who is now 6 months old) with the intention of doing computational analysis of the development of his speech patterns.

I haven't much deep experience with ML/stats and was looking for some guidance as to what kind of signal processing (to clear out background noise, etc) as well as statistics procedures (clustering, cross-similarity of clusters across datasets, etc) I could run on my data and what kind of insight I would get from this.

  • 4
    $\begingroup$ You're basically asking how to do speech analysis and computational linguistics, both of which are enormous fields $\endgroup$
    – Alex R.
    Apr 8 '16 at 0:30
  • $\begingroup$ Yup. But looking for the gist of it so I can at least do something. I don't always need to make things as complex as possible – just enough to advance my understanding so that one day I can study this stuff in all its rigor. $\endgroup$ Apr 8 '16 at 0:34

This is a pretty broad question, but I've been working with general speech datasets, so I can try give an idea of what could be done, and share some things I would probably try myself.

Normally, in adult speech recognition, what is asked that people read a certain text, but this is obviously impossible with babies. So instead, I would listen to the speech files and look for similar speech samples, which should give an idea of what could be done.

Linguistic analysis

I think it is pretty difficult to do anything from a linguistical perspective. At this stage, vowel formation is happening, so running some vowel recognisers might be an interesting thing to try. I believe there are some which output the probability of which vowel was formed and then you could do a longitudinal analysis of increase in vowel formation quality (based on the recognition probabilities).

Spectral analysis

I would probably look at the spectral components of the speech, i.e take a mean spectogram at each age stage and compare. I would also try Mel Frequency Cepstral Coefficent (MFCC) representation, as that is a more perceptual representation of speech more commonly used.

What has been done

As I know, child speech is an active research area, I'm familiar that there is a research group in Lisboa working on child speech and child speech therapy, see Martins et al. : Detection of Children's Voices, but "toddler speech" is a bit more difficult. One of their member's, Alberto Abad, gave a presentation here, also from the perspective of pathological child speech and development, which seems especially difficult. I found that interesting, I recommend you to watch the beginning so that you can have an idea what speech structures even exist at that stage.

The tools that you will need to undertake the analysis really depends on the choice of your programming language (Python or R), but I hope I gave you some initial directions.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.