I recently started to work on Speaker/Language recogntion using i-vector, and after consluting with researcher on researchgate, I came to the following steps:
1) Database
i) Developement dataset (UBM, T training), if labeled (LDA and PLDA also)
ii) Training dataset(For speaker/Language Enrollment, modeled speakers), if the Developement dataset is not labeled, I trained LDA and PLDA on training dataset(needs comments on this)
iii) Testing dataset (for testing the modeled speakers/language)
About Language Detection:
If I have lot of speech samples, but no labled for that speech utterance, how can I train LDA/PLDA for languages? or can I trained these on training languages data?
What about the Gender? how much the results will be effected if we have different/same UBM, T? Is it ok to have single UBM, T for both genders?
Is there any way to apply the i-vector detection without applying LDA and PLDA such as SVM on i-vectors without i-vector reduction??
I applied nonparametric Bayesian mixture model ( known as DPM) few years ago to the the problem of language identification (using i-vectors). The results was very interesting since in addition to automatic clustering it was also showing relationship between languages (e.g. which language is similar to other etc). unfortunately, the project was not owned by me ( I did the experiment for someone else using my code) and the institution that owned the data did not want to publish the results. However, you can find more about the DPM by googling. I also have a C++ implementation (I used this for the above ) that you can find among my publications.
@Amir: Its depend how I can apply, is there any report which I can use to uderstand your concept, or you can mentions steps??? and It would be great if I can have your C++ implementation.
@Amir are you talking about
http://scikit-learn.org/stable/modules/generated/sklearn.mixture.DPGMM.html
http://scikit-learn.org/stable/modules/generated/sklearn.mixture.DPGMM.html
I have not checked that code. I have my own c++ based: https://www.researchgate.net/publication/289035262_hdphmm_lib
Code hdphmm lib
You don't have to perform i-vector reduction using PLDA/LDA, in that case you can use Cosine Distance Scoring for comparison with your model i-vectors.
I was using i-vectors for age estimation task. I used Within Class Covariance Normalization as channel compensation technique, where the classes were speakers (as suggested in this paper: http://www.sciencedirect.com/science/article/pii/S0952197614001018). You can always try i-vector approach without session compensation and then compare those results with i-vectors+WCCN approach.
- Badges
- Science topic