I'm reading a thesis, it mainly uses a GMM-HMM to tackle the acoustic unit discovery task. It proposes a delicate Bayesian framework on top of the GMM-HMM. In page 15 it says

Notice that we have not included any parameters of the transition probabilities as it has been empirically observed that they play no significant role when modeling speech (Bourlard, 1996). Consequently, we assume the transition probabilities are fixed parameters such that the probability to go to any state given the current state is the same.

I guess the claim is something general (so the Bayesian things won't matter much). However, the reference (Bourlard, 1996) seems to be an old French literature that I cannot access. Can someone explain briefly why it's true or point me somewhere that is accessible?

Thank you in advance.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.