I have audio and its spectrogram of the words in English language. (A spectrogram is a frequency domain representation of a signal) Consider the words: chain, change, chair, chapter. As you can notice, the 'ch' sound is common among the words considered.

Is there an algorithm that will allow me to identify which part of the spectrogram (data) is common between the spectrograms of the words? In other words, is there an algorithm that can identify the part of the spectrogram that represents 'ch' sound in each of the words?

  • $\begingroup$ In case anybody is wondering, I have decided to use Locality Sensitive Hashing (LSH) to hash every frame of the spectrogram. All frames from different spectrograms that have the same hash value, mostly likely belongs to a phoneme that is common to the words of those spectrograms. This technique isn't perfect, but surprisingly works quite well, when a probability is assigned to every phoneme based on the output of LSH. $\endgroup$ – Vijayaraghavan Purush Sep 18 '19 at 4:09

The "ch" sound is called a phoneme. Automatic speech segmentation on the phoneme level (as opposed to word or sentence) will allow you to extract each phoneme from the speech samples. This is sometimes referred to as phoneme segmentation. Several papers for this are available in the literature.

Once you have the phonemes extracted, you can compute distances/similarities between them (in some feature space). An example feature representation could be MFCC.

| improve this answer | |
  • $\begingroup$ Thanks for the term. Can you cite few phoneme segmentation techniques or papers? $\endgroup$ – Vijayaraghavan Purush Sep 9 '19 at 9:17

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.