A parameter estimation algorithm for speech recognition to maximize 'state optimized joint likelihood'