2
$\begingroup$

I'm trying to implement a speaker recognition system and want to make sure I'm aware of the latest trends in speech segmentation. I've read a number of very different methods at a higher level but I'm not sure which one would be best for my particular situation. I'll list what I found so far and I'd appreciate either (a) comments on the research I've done so far or (b) popular methods that I didn't mention that I should know about. Then I'll take it from there in terms of determining what will work best for my situation.

Techniques for Speech Segmentation (that I've identified so far)

  • I've read that strong long-term modulation frequencies between 1-16 Hz are indicative of speech activity, but haven't been able to find any good explanations of what exactly these are. What I've been able to determine so far is that they are the temporal spectrum of specific frequency bands. (Unsupervised speech/non-speech detection for automatic speech recognition in meeting rooms by Maganti et al)

  • Using energy variance (high variance implies speech-like segments) independently or with other methods to build a Gaussian mixture model with two components: one for speech and one for noise, and then re-classifying windows as either speech or non-speech. (An unsupervised, sequential learning algorithm for the segmentation of speech waveforms with multiple speakers by Siu et al)

  • Using the Voting Experts algorithm, which utilizes iter- and intra- window entropy (a window in this sense is a 'window of windows') to determine where "chunks" are using a feature that has high inter-segment entropy, and then to classify these chunks as either speech or non-speech using 2-means clustering, where each cluster of chunks contains all chunks of one of the two classes. (Voting Experts: an unsupervised algorithm for segmenting sequences Cohen et al)

The short version of my question

Ultimately, I'd really like to know what is standard in the industry as of right now. Then I can determine what will work best for my particular situation once I have a little more base knowledge.

| improve this question | |
$\endgroup$
  • $\begingroup$ Could you add some references to your "identified techniques" section? Also, what exactly is your particular situation? $\endgroup$ – user13107 Jul 20 '13 at 4:42
1
$\begingroup$

I'm going to answer my own question, for those who are looking for the same information as I was. Adaptive Thresholding is the most common way to segment audio, and it gives you the most bang for the buck. While there are more accurate methods out there for segmenting, adaptive thresholding is fast and very accurate, and uses simple features such as zero-crossings, energy, and entropy. I'm going to have to look into this more myself but this should be enough to get started.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Can you provide links/references? $\endgroup$ – endolith Jul 25 '13 at 21:50
  • $\begingroup$ The references are listed about- you should be able to find all of them with a quick google $\endgroup$ – wheresmycookie Jul 25 '13 at 23:35

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.