While I can't give you a reference, I can explain the difference in general terms.

It all depends on which features you use.

Since HMM is a generative model, it is hard to make it use a large number of features. SVM is a discriminative model so it can handle large feature sets very well.

Thus, with the same features, it is unclear whether HMM or SVM is better (probably HMM since generative models often generalize better). But SVM allows you to incorporate more features that help you attain a better performance.

So, without having experience in emotion detection, I would recommend a discriminative approach if you want to get at high levels of accuracy.

Dear Gabriele,

What I would say is we have to try both for finding out that as it totally depends on the type and scale of data and its features. Another point we can say that HMM is a bit risky while SVM is more secured generally. 

Thanks

Hi Gabriele,

In my experience, there's no "best" method in this area. Each method has its own advantages and will perform differently in every researches.

I suggest you to compare them by yourself.

Cheers,

Diky

Hello. It is not clear to me that SVM and HMM solve the same statistical estimation problem. For me, SVM is essentially a classifier, i.e., it estimates the decision boundary between different classes. The HMM framework fits a Markov model to the data, which can be used to estimate the most likely sequence of states via one of two optimality criteria (instantaneously most probable = Forward backward, or most probable sequence = Viterbi algorithm). I would feel more comfortable comparing the performance of SVM with a clustering technique like Gaussian mixture models or a supervised approach like decision trees. Naturally, your results will depend on your data.