A method and system for indoor early reflection sound localization

Abstract

The invention discloses an indoor early reflection sound positioning method and system. The method is as follows: 1) generating impulse responses of different rooms; 2) for each impulse response, performing convolution calculation with the sound source signal to obtain an array signal of M channels; 3) performing a calculation on each array signal Short-time Fourier transform, and take J time-frequency points; then calculate the continuous N-order HOA coefficients of each time-frequency point, and convert the array signal of M channels into N 2 channel array signal; 4) each N 2 The real part and imaginary part of the channel array signal are spliced ​​separately as separate channels to obtain a 2N 2 The array signal of the channel; then pack the continuous K frames to form a 2N 2 ×K×J dimension samples; 5) use the samples to train the neural network; for a sound source in a target room, the impulse response of the target room and the corresponding 2N 2 The ×K×J dimensional array signal is input into the trained neural network to locate the position of the sound source in the target room.

Description

technical field [0001] The invention belongs to the technical field of sound source localization, and in particular relates to a method for predicting the arrival directions of direct sound and indoor early reflection sound based on a deep residual network. Background technique [0002] When the sound source is sounding indoors, reverberation (reverberation) will occur due to reflections from walls, ceilings, and floors. Related studies have shown that the early reflected sound components in reverberation (arriving at the receiving point within 50ms later than the direct sound) are helpful to improve speech intelligibility. Early reflections are signals related to the intensity of the direct sound and contain most of the information in the direct sound. If the azimuth information of the early reflected sound is known, signal extraction can be performed for each direction beamforming, and then the purpose of signal enhancement can be achieved. On the other hand, early refle...

AI Technical Summary

Problems solved by technology

Tervo et al. used SRP-PHAT and GCC-PHAT algorithms for early reflection positioning, but the performance of both of them dropped sharply under strong reverberation, and the latter also had the disadvantage of error accumulation

Tervo also proposed a reflection path tracing method based on the measurement of room impulse response, but the actual operation process of this method is not easy to implement and does not have generalization

The EBMVDR method proposed by Sun et al. transfers beamforming from the traditional spatial domain to the characteristic beam domain (spherical harmonic domain), avoiding the singular matrix problem caused by coherent signals, but manual setting and adjustment are required in the implementation process Focus on frequency, and can only use a narrow frequency smoothing range, it is difficult to achieve more robust anti-noise performance

Images