Are you an EPFL student looking for a semester project?
Work with us on data science and visualisation projects, and deploy your project as an app on top of Graph Search.
Recently, a new method for sub-band based ASR has been introduced which approximates the ideal 'full combination' (FC) approach. We show how this ideal approach can be set up as a nonlinear combination function and how it can be approximated by estimating the posteriors for each combination as a function of the posteriors from the individual sub-bands alone [Hagen98:SBS,Morris99:TFC]. Sub-band based ASR is especially promising in stationary and time-varying band-limited noise as long as the remaining clean sub-bands supply sufficiently reliable information. For the first time, experiments will be presented which compare the FC approach and its approximation on these noise characteristics. They were carried out on PLP [Hermansky90:PLP] and J-RASTA-PLP [Hermansky94:RPO] features in the framework of HMM/ANN Hybrid systems.
Rolf Gruetter, Ileana Ozana Jelescu, Ting Yin, Yujian Diao
Mihai Adrian Ionescu, Junrui Zhang, Francesco Bellando, Pierpaolo Palestri, Luca Selmi