Clients have still hesitated to switch conventional phone service with voice over IP networks (VoIP) service because VoIP service providers are not successful in providing consistent quality during a call. The uncertainness of IP networks, the legacy of packet-switched networks, makes it hard to predict service quality and demands real-time based monitoring. In this paper, we propose a prediction voice quality metric to monitor the quality of VoIP service. Based on a learning machine, the proposed metric nonlinearly weighs network parameters to estimate speech quality. Finally, performance analysis shows that the proposed metric achieves the high prediction accuracy.