Tone Quality Improvement of Bone Conduction Voice by Cepstrum-based Local Conversion Models

A novel tone quality improvement method for a bone conduction voice is presented. In the present method, the tone quality of the bone conduction voice is converted to the similar quality of the air conduction voice. For the voice conversion, the present method uses a codebook, which consists of various paired code vectors of the bone and air conduction voices. The deltaand mel-cepstral coefficients are employed as the code vectors. The delta-cepstral coefficients in the code vectors are first quantized and classified by a neural-gas’ network. The relationship between the mel-cepstral coefficients of the bone and air conduction voices in each class is described locally by a mathematical conversion model. The bone conduction voice is then converted into the clear air conduction voice by using those mathematical local models. The validity and effectiveness of the present method have been confirmed by applying it to the tone quality conversion problem of the real bone conduction voice. Key–Words: Bone conduction voice, Delta-cepstrum, Neural-gas’ network, Local conversion model