Noise Estimation Method and System

Abstract

The invention discloses a noise estimation method and system, wherein the noise estimation method comprises the steps of: receiving noisy voice data; extracting an original noise signal in the noisy voice data; and compressing the original noise signal according to a frequency division standard to obtain a target noise signal. By using the noise estimation method, the subsequent computation burdencan be effectively reduced on the basis of a finite hardware resource, and meanwhile, performance loss can be avoided. Further, accurate noisy data can be obtained on the basis of a classifier trained by mass data, so that the reliable noise prediction basis can be provided for subsequent relevant operation of reducing the computation burden. Furthermore, unstable mutant noise can also be effectively traced according to the prior knowledge of a scene noise characteristic and the accurately extracted noise signal.

Description

technical field [0001] The present invention relates to the field of signal processing, in particular to a noise estimation method and system. Background technique [0002] Speech recognition technology is widely used, such as in-vehicle voice interaction systems, etc. In order to ensure the accuracy of recognition, it is necessary to reliably predict the quality of audio input and process audio that may interfere with the subsequent recognition process. For example, the voice data collected in a moving car will contain a lot of background noise, which will easily cause interference to related voice recognition, action execution, and communication quality. Therefore, it is usually necessary to perform noise reduction processing on the received audio input signal. [0003] For the estimation of the speech signal, an estimator based on the logarithmic magnitude spectrum is generally used. Under the model assumption of the Gaussian signal, the mean square error of the logarithm...

AI Technical Summary

Problems solved by technology

[0004] Traditional noise estimation methods mostly use speech activity detection as the basis for discrimination, and use statistical information to predict noise in the noise segment, but the ability to track sudden noise is poor, and in the case of low signal-to-noise ratio, the noise segment and The distinction between speech segments is not obvious, and effective noise estimation cannot be performed

In particular, due to the cost considerations of the entire industry, under the premise of limiting the number of hardware and specifications (using a single microphone for noise reduction calculations, and using smaller ROM and RAM chips), in order to obtain good noise reduction processing As a result, the requirement for computational complexity in the noise reduction process is high, and the previous noise estimation process has a direct impact on the computational complexity of the noise reduction process.

Images