Deep Neural Network Voice Direction Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current direction of arrival (DOA) estimation and voice activity detection methods face challenges in accurately determining the presence of sub-voices in specific direction intervals, leading to incomplete or inaccurate predictions.

Innovation Solution

A method utilizing a pre-trained deep neural network, comprising a combination of convolutional neural networks, feature-extraction networks, and Fourier transform networks, to predict the presence of sub-voices in each direction interval by inputting target voices and processing them through multi-head fully connected networks, feature-extraction layers, and convolutional neural layers to generate accurate prediction results.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional DOA estimation methods using receiving apparatus arrays are used, then the direction of arrival can be determined through triangulation, but the accuracy of predicting sub-voices in specific direction intervals is insufficient

Engineering Contradiction:
Improvedirection of arrival estimation accuracyVSAvoidsub-voice detection accuracy
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent divides the 360-degree direction space into multiple direction intervals (e.g., 0-30°, 30-60°, etc.) and uses separate neural network prediction heads for each interval. This segmentation allows the system to independently analyze and predict sub-voice presence in each directional sector, improving both DOA estimation accuracy and sub-voice detection reliability by focusing computational attention on specific angular regions rather than treating all directions uniformly.

Inventive Principle:
Principle #1Segmentation

2Extent of automation

If convolutional neural networks are used for DOA estimation as described in prior art, then automated voice detection is improved, but the ability to separately predict and accurately identify sub-voices in each direction interval remains limited

Engineering Contradiction:
Improveautomated voice detectionVSAvoiddirectional sub-voice prediction accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent extends the neural network architecture from a single-output design to a multi-dimensional output structure where each direction interval has its own prediction head. This dimensional expansion transforms the problem from predicting a single DOA value to simultaneously predicting sub-voice presence across multiple directional dimensions, enabling automated detection while maintaining high directional prediction accuracy through parallel processing of angular information.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP3816999B1Neural network based determination of the direction of arrival of voice
Publication Date: 2022.11.09 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP3816999B1 patent drawingFigure 1
  • EP3816999B1 patent drawingFigure 2~3A
  • EP3816999B1 patent drawingFigure 3B

AI summary

The present disclosure provides a method and apparatus for detecting a voice, relates to the fields of voice processing and deep learning technology. A specific embodiment includes: acquiring a target voice; and inputting the target voice into a pre-trained deep neural network to obtain whether the target voice has a sub-voice in each of a plurality of preset direction intervals, the deep neural network being used to predict whether the voice has a sub-voice in each of the plurality of direction intervals. The present disclosure may predict each direction interval separately, so as to accurately determine whether the target voice has a sub-voice in each direction interval, to realize accurate prediction.