Interspeech'2005 - Eurospeech

Lisbon, Portugal
September 4-8, 2005

Extended Baum-Welch Reestimation of Gaussian Mixture Models Based on Reverse Jensen Inequality

Mohamed Afify

BBN Technologies, Cambridge, MA, USA

In this paper we derive the well known EBW reestimation formulae for Gaussian mixture models using the recently proposed reverse Jensen inequality. In addition to the simplicity of the derivation, it leads to closed form expressions for the D of each Gaussian in the mixture. Using some approximations, it is shown that the expressions can be reduced to the popular formula of [8] with a Gaussian dependent step size. The average of the normalized distance used in calculating the step size is empirically verified to be very close to one. Hence, the proposed formula leads to a global value very close to 3. The obtained results are validated in experiments on large vocabulary speech recognition.

Full Paper

Bibliographic reference.  Afify, Mohamed (2005): "Extended baum-welch reestimation of Gaussian mixture models based on reverse Jensen inequality", In INTERSPEECH-2005, 1113-1116.