Skip to content

Latest commit

 

History

History
124 lines (79 loc) · 4.17 KB

CRNN-Based Multiple DoA Estimation Using Acoustic Intensity Features for Ambisonics Recordings阅读笔记.md

File metadata and controls

124 lines (79 loc) · 4.17 KB

#! https://zhuanlan.zhihu.com/p/447506072

CRNN-Based Multiple DoA Estimation Using Acoustic Intensity Features for Ambisonics Recordings阅读笔记

Introduction

CRNN to estimate the DOAs from a first-order Ambisonics (FOA) recording.

Input: features derived from the acoustic intensity vector

Consider a normalized expression of the acoustic intensity vector in each time-frequency bin and propose to use its coefficients as input features.

Background

A. Ambisonics Format

The sound field is recorded by a spherical microphone array and converted into Ambisonics with an encoding matrix.

FOA corresponds to the coeff of the decomposition in the spherical harmonics of order 0 (channel W) and 1 (channels X, Y and Z) $$ \left[\begin{array}{c} W(t, f) \ X(t, f) \ Y(t, f) \ Z(t, f) \end{array}\right]=\left[\begin{array}{c} 1 \ \sqrt{3} \cos \theta \cos \phi \ \sqrt{3} \sin \theta \cos \phi \ \sqrt{3} \sin \phi \end{array}\right] p(t, f) $$ 在 $p(t,f)$ 处的 FOA 表示。

B. Acoustic Intensity

The active intensity vector $I_a(t,f) = \mathcal{R}{p(t,f)v^*(t,f)}$ represents the flow of sound energy in a point of space, with $v(t,f)$ the particle velocity. $$ \mathbf{v}(t, f)=-\frac{1}{\rho_{0} c \sqrt{3}}\left[\begin{array}{c} X(t, f) \ Y(t, f) \ Z(t, f) \end{array}\right] $$ $\rho_0$: the density of air; $p(t,f)=W(t,f)$

The active intensity vector disregarding the constant: $$ \mathbf{I}{\mathrm{a}}(t, f)=-\left[\begin{array}{l} \mathcal{R}\left{W(t, f) X^{}(t, f)\right} \ \mathcal{R}\left{W(t, f) Y^{}(t, f)\right} \ \mathcal{R}\left{W(t, f) Z^{}(t, f)\right} \end{array}\right] $$ The reactive intensity vector $I_r(t,f) = \mathcal{I}{p(t,f)v^(t,f)}$, represents dissipative local energy transfers. $$ \mathbf{I}{\mathrm{r}}(t, f)=-\left[\begin{array}{l} \mathcal{I}\left{W(t, f) X^{}(t, f)\right} \ \mathcal{I}\left{W(t, f) Y^{}(t, f)\right} \ \mathcal{I}\left{W(t, f) Z^{*}(t, f)\right} \end{array}\right] $$

DOA Estimation System

A. Input Features

Propose to exploit both the active and reactive intensity vectors across all freq bins in the STFT domain as inputs to the neural network in a given time frame. Motivated by the fact that the active intensity relates more directly to the DOA and the reactive intensity indicates whether a given time-freq bin is dominated by direct sound from a single source, as opposed to overlapping sources or reverberation.

normalize the inputs in each tf bin regardless of the sound power: $$ \frac{-1}{C(t,f)}\left[\begin{array}{c} \mathbf{I}{\mathrm{a}}(t, f) \ \mathbf{I}{\mathrm{r}}(t, f) \end{array}\right] $$ B. Target Outputs and Training Cost

The target output of the CRNN is a binary vector of size $n_{DOA} \times 1$, each index corresponds to one discrete DOA. The element of the target vector that is the closest to the true DOA is set to 1. (>1 elements can be set to 1 when multi-sources)

Train a specific neural network for each number of sources.

C. Network Architecture

image-20211214104951690

T (num of frames): 25, F (num of freq bins): 513, C (num of feature channels): 6

Convolutional modules aim to extract spatial information from the inputs. (Convolve along freq)

The second part (2 BiLSTM and 2 FC) uses this information to estimate the DOAs.

D. From Framewise to Global DOA Estimation

Analysis by LRP

Layer-wise Relevance Propagation (LRP) is a technique for determining which features in a particular input vector contribute most strongly to a neural network’s output.

Shared Experimental Settings for DOA Estimation

16kHz

STFT: win1024, hop 512

B. Training Procedure

Each network could be used to predict any number of sources, but training each network for a specific number of sources yielded better results.

neighborhood of the peak: $\Delta=2\alpha$, $\alpha$ is the angular resolution.

Nadam optimizer, initial lr 10e-3, 0.2 for the single-source network/0.3 for the two-source network dropout after conv block, FC and on the recurrent weights of the BiLSTM layers

early stopping with a patience of 20 epochs. 80/150 epochs for the single-source network and the two-source network.