A Studyon an LP-based Model forRestoring Bone-conducted Speech

Inahighly noisy environment, bone-conducted speech seemstobe more advantageous thannormalnoisyspeech because ofitsstability against surrounding noise. Thesound quality ofbone-conducted speech, however, isverylowand restoring bone-conducted speechisa challenging newtopic in speechsignal processing field. Inthispaper,we proposea restoration modelbasedonlinear prediction (LP). Toevaluate theability oftheLP-based modeltoimprove thevoice quality, we compared itwithexisting modelsusing onesubjective andthree objective measurements. Theexperiments showedthattheLP- basedmodelyields restored signals thatarebetter forboth humanhearing andforthefront-ends ofautomatic speech recognition systems. Astherestoration ability oftheLP-based modeldependedon a fewparameters related to theLP coefficients ofair-conducted speech, we applied a multi-layer perceptron neuralnetworkto blindly predict them with reasonable results.