An approach to noise cancellation for speech recognition systems

Abstract

This paper describes an approach to simple noise cancellation for speech recognition, such as template matching on time‐spectrum patterns. This system has two inputs. The primary input Ip receives a speech signal S(t) corrupted by an environmental noise N(t). The second input Ir, used as a reference signal, receives kN(t + td), that is, a correlated version of the noise in Ip, where k represents the level ratio of kN(t + td) to N(t) included in the primary signal, and td is the time difference between Ip and Ir. The two input signals are analyzed using a bandpass filter bank (BPFB). The noise cancellation is performed in the following manner: (1) The level ratio k is estimated; (2) the time difference td is also estimated using k; (3) the cancellation is performed on every output of the BPFB using k and td. At the same time, a speech section is detected by using the variations in each spectrum difference between the two input signals. The values of the parameters k and td are fixed at the beginning of the speech section while a speech signal exists.