Bachelor Theses - Details

Recently, processing of body-conducted speech signals has gained interest, as accelerometer-equipped earbuds offer easy access to these signals. These accelerometers capture the user's own voice relatively directly and without outside noise, promising improved speech enhancement, among other applications.

While much research exists on the processing of body-conducted speech, the exact characteristics of the bone-conduction of speech are often overlooked. These characteristics can vary not only depending on the person, but also on the currently spoken phoneme and pitch. Especially how much the body-conduction actually varies is of great interest, because it has implications on the possible performance of systems processing body-conducted speech. The goal of this thesis is to analyze the properties of body-conduction of speech, and to derive specific transfer functions if possible.

First, the literature shall be reviewed for relevant existing research. Then, measurements shall be conducted to characterize the transmission. For this, one or multiple accelerometers shall be fixed to an earbud and possibly other positions on the head. The accelerometers as well as other microphones shall be recorded using speech and different phonemes as excitation. Afterwards a thorough analysis of the resulting dataset shall be conducted. The relation of the recorded signals shall be characterized, and, if possible, transfer functions shall be derived. Of special interest is identifying clusters based on phoneme type, as well as the variance of the transfer functions depending on fit. If possible, measurements and analysis shall be repeated with different subjects.