Multi-microphone neural network for sound recognition

Abstract

A neural network is provided for recognition and enhancement of multi-channel sound signals received by multiple microphones, which need not be aligned in a linear array in a given environment. Directions and distances of sound sources may also be detected by the neural network without the need for a beamformer connected to the microphones. The neural network may be trained by knowledge gained from free-field array impulse responses obtained in an anechoic chamber, array impulse responses that model simulated environments of different reverberation times, and array impulse responses obtained in actual environments.

Description

BACKGROUND[0001]Multi-microphone arrays have been used to enhance sound detection. Traditional schemes of multi-microphone sound detection include using a beamformer to combine sound signals received from multiple microphones to form an enhanced single-channel audio signal, which may be transmitted to a sound recognition system. The beamformer and the recognition system are typically designed independently of one another. As a consequence, enhancement of audio signals by the beamformer may not be optimal for the recognition system. Moreover, such beamformers typically require “tractable” array geometries such that the scattering and diffraction effects and beamformer response can be easily analyzed or optimized. However, such tractable array geometries may limit the number of design options for multi-microphone devices. For example, in conventional beamforming schemes which require tractable array geometries, the user may have a limited number of options in the physical placement of...

AI Technical Summary

Benefits of technology

The patent describes a method to train a neural network for multi-microphone sound recognition. This involves generating free-field array impulse responses over multiple directions based on sound events received by multiple microphones. A first set of array impulse responses is generated to model simulated environments of different reverberation times, while a second set of array impulse responses is generated in actual environments. The neural network is then trained based on these impulse responses to recognize different sound patterns. The technical effects of this method include improved sound recognition accuracy and improved training efficiency.

Problems solved by technology

As a consequence, enhancement of audio signals by the beamformer may not be optimal for the recognition system.

However, such tractable array geometries may limit the number of design options for multi-microphone devices.

For example, in conventional beamforming schemes which require tractable array geometries, the user may have a limited number of options in the physical placement of microphones in a given environment.

Images