Real-Time Labeling of Talker Identity Using Short Speech Segments in a Microphone-Array Environment

The problem addressed is generally in the family of speaker verification, but its conditions and requirement s are very different from what has been published in this area. We want to label approximately five to ten moving talkers in the reverberant and noisy focal area of a large-aperture microphone array, in real time, from short segments of conversational audio. Current real-time systems use only spatial informat ion, which is inadequate when talkers move while silent. Given th e dynamic noise conditions in this kind of environment, it is very difficult to collect sufficient training data for a conventional algorithm. The proposed algorithm is easily implementablein real time and is trained offline using only noise-free data. This is possible because a set of robust features is used that represent the differences between a pair of speech segments . Also, the output of the algorithm is a probability, rather th an a measure requiring threshold tuning, which makes on-the-fly decisions straightforward. The algorithm is introduced and its properties are demonstrated. While quite simple, the algor ithm has been shown to outperform some commonly-used talkerdistance metrics for real data taken from a noisy-environment, array system. This microphone-array dataset is available t o other researchers at www.lems.brown.edu/array/data/movingta lkers/. EDICS Categories: SPE-SPKR, AUD-LMAP