| One of the principle challenges in the area of speaker identification is enabling currently available applications to work in operational environments. In an operational environment speaker identification systems are confounding with reverberation and other channel effects; several sources of noise and non-stationary interference; and a multitude of speech like sounds, such as coughing and laughter, which lie outside the capabilities of even the most robust speaker identification systems. To address these issues, the notion of usable speech was developed. As opposed to the filtering paradigm, the usable speech concept is to identify and extract those segments of signal that are known to perform well with the intended application. A particularly difficult and common type of non-stationary interference will be dealt with in this research, which is that of an inferring talker. This type of situation is called co-channel speech. However, it should be noted that the techniques developed in this research can be applied to many situations where there exists both stationary and non-stationary interference. The co-channel usable speech concept will be further explored by the detailed statistical analysis and characterization of usable and unusable speech segments using a continuous-time Markov chain model of the voiced, unvoiced, and silent regions of the speech signal. In addition, a novel structure-based usable speech feature is developed that takes into consideration the deleterious effects of jitter, shimmer, and diplophonia. This feature is then used to derive a context dependent estimator of the instantaneous interference level of the corrupted speech signal, which can be used to detect the less degraded portions of the signal. In addition, three novel applications of the usable speech concept will be developed: co-channel speaker identification, co-channel speaker separation, and a simple speaker counting system. |
