Voice Detection Model Using RNN for Noisy Environments

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional voice activity detection algorithms face challenges in accuracy and stability, especially in noisy environments and real-time conference communication devices, due to dependence on signal-to-noise ratio and large parameter scales of deep learning models.

Innovation Solution

A method involving mixed voice synthesis, 62-dimensional feature extraction, and training with an RNN model using an improved loss function and Adam training strategy to develop a compact and stable voice detection model, capable of recognizing voice data in various noise conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If traditional VAD algorithms are used, then the algorithm is simple to implement, but the accuracy of voice detection in noisy environments deteriorates

Engineering Contradiction:
Improvealgorithm complexityVSAvoidvoice detection accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent introduces an intermediary deep learning model (RNN/LSTM/GRU) that bridges the gap between simple traditional algorithms and complex accurate detection. This intermediary model processes acoustic features and generates voice activity predictions, achieving high accuracy in noisy environments while maintaining reasonable computational complexity through optimized architecture design.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms the detection approach by changing parameters from traditional fixed thresholds to dynamic neural network predictions. The system uses multiple acoustic features (energy, spectral centroid, zero-crossing rate, etc.) as input parameters to the deep learning model, which learns optimal parameter combinations for accurate voice detection in varying noise conditions.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If deep learning models (RNN, LSTM, GRU) are used for VAD, then the voice detection accuracy improves, but the parameter scale and calculation overhead increase

Engineering Contradiction:
Improvevoice detection accuracyVSAvoidmodel parameter scale
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the deep learning model into distinct functional components: feature extraction layer, hidden layers (RNN/LSTM/GRU), and output layer. This segmentation allows for optimized training of each component and enables deployment on devices with limited resources by processing audio in frames and using only necessary model parts.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements partial action by processing audio in overlapping frames rather than analyzing entire audio streams at once. The model uses a subset of acoustic features and processes data in manageable chunks, achieving accurate voice detection without requiring the full computational power that would be needed for exhaustive analysis of all audio parameters simultaneously.

Inventive Principle:
Principle #16Partial or excessive action

3Use of energy by moving object

If traditional VAD algorithms with fixed thresholds are used, then the calculation overhead is low, but the stability and reliability in real-time communication deteriorate

Engineering Contradiction:
Improvecalculation overheadVSAvoidalgorithm stability
Core Design Contradiction:
Use of energy by moving objectVSReliability

Solution Approach 1:

The patent applies preliminary action by pre-training the deep learning model on diverse audio data including various noise conditions before deployment. The model learns optimal detection patterns in advance, enabling it to maintain high reliability in real-time communication without requiring complex runtime adjustments or recalibration, thus balancing computational efficiency with stable performance.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3739582B1Voice detection
Publication Date: 2023.07.26 YEALINK (XIAMEN) NETWORK TECHNOLOGY CO LTD
  • EP3739582B1 patent drawingFigure 1~2
  • EP3739582B1 patent drawingFigure 3~4
  • EP3739582B1 patent drawingFigure 5~7

AI summary

The present disclosure discloses a method for constructing a voice detection model and a voice endpoint detection system, and belongs to the technical field of voice recognition. In the method for constructing a voice detection model according to the present disclosure, audio data is first collected and a mixed voice is synthesized, feature extraction is performed on the mixed voice to obtain a 62-dimensional feature, and then the 62-dimensional feature is input to a recurrent neural network (RNN) model for training to obtain a voice detection model. The voice endpoint detection system according to the present disclosure includes a collecting unit, a calculating unit, a transmitting unit, and a terminal, the collecting unit being electrically connected to the calculating unit, and the calculating unit and the terminal being respectively connected to the transmitting unit. In order to overcome the shortcomings of low accuracy and stability of voice recognition in the prior art, the present disclosure provides the method for constructing a voice detection model and the voice endpoint detection system, so that voice data in the case of various types of noise has a high recognition rate, and the voice detection model can be applied to a real-time conference communication device.