Identifying the number of sources in speech mixtures with the Mean Shift algorithm

Blind Source Separation (BSS) of speech sources has been subject of study during many years, and it still remains largely open and unsolved. Traditional BSS methods based on statistical properties of the signals, as well as recent methods such as time-frequency masking, normally need to know in advance the number of sources in the mixture to perform the separation. Additionally, there are many applications where speech source enumeration can be very useful. Bearing in mind the need of automatic source enumeration of speech sources, this paper proposes a novel pruning algorithm based on the mean shift algorithm for speech separation. The results obtained identifying the number of sources in linear mixtures of 2, 3 and 4 sources support the suitability of the proposed algorithm. Key–Words: Blind Source Separation, Source Enumeration, Speech enhancement