Genetic algorithm-based improvement of robot hearing capabilities in separating and recognizing simultaneous speech signals

Since a robot usually hears a mixture of sounds, in particular, simultaneous speech signals, it should be able to localize, separate, and recognize each speech signal. Since separated speech signals suffer from spectral distortion, normal automatic speech recognition (ASR) may fail in recognizing such distorted speech signals. Yamamoto et al. proposed using the Missing Feature Theory to mask corrupt features in ASR, and developed the automatic missing-feature-mask generation (AMG) system by using information obtained by sound source separation (SSS). Our evaluations of recognition performance of the system indicate possibilities for improving it by optimizing many of its parameters. We used genetic algorithms to optimize these parameters. Each chromosome consists of a set of parameters for SSS and AMG, and each chromosome is evaluated by recognition rate of separated sounds. We obtained an optimized sets of parameters for each distance (from 50 cm to 250 cm by 50 cm) and direction (30, 60, and 90 degree intervals) for two simultaneous speech signals. The average isolated word recognition rates ranged from 84.9% to 94.7%.