Evaluation of binaural Own Voice Detection (OVD) algorithms

This paper evaluates three different two-channel algorithms developed to detect the presence of a speaker's own voice, if the recording microphones are placed bilaterally close to the ears. In this evaluation, a typical conversation is simulated with a spatially diffuse cafeteria noise as a background noise source. The results show that none of the algorithms can separate short signal blocks with or without the speaker's own voice in any circumstance. Especially when the background noise is very quiet or very loud, the estimation seems to be difficult. In a second experiment, we show that the algorithms will perform better, if the conversation partner is not exactly in front of the speaker. The comparison shows that the coherence-based algorithm is the best choice for further development.