Spectral Subtraction Method utilizing Asynchronous Signals Observed over TCP/IP based Network

摘要

An utterance training system utilizing automatic speech recognition has been developed as computer aided language laboratory system. Many noise reduction methods are proposed because a performance of speech recognition is seriously degraded by surround noise. Spectral subtraction, one of the noise reduction methods, suppresses stationary noise from a signal by subtracting a noise spectrum observed in a signal-free duration. When an observed noise spectrum includes spectral components of a target signal, the performance of spectral subtraction would be degraded. In particular, the lower the signal to noise ratio, the larger the degradation. In this paper, a spectral subtraction method using signals observed at separate computers connected by network is proposed. Because of TCP/IP based network characteristics, the transmission delays among computers are varied due to various reasons such as collision of packets. The proposed method assumes that the noise spectrum is stable for certain duration and utilizes it for spectral subtraction in order to overcome the un.xed transmission delays. Simulation results show a possibility to improve the performance of speech recognition when several interference speakers exist around the target speaker.