CNN Sound Signal Model Learning for Stable Source Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep clustering methods using bidirectional long short-term memory (BLSTM) networks face instability and overlearning issues, leading to inefficient learning and sound source separation in acoustic scene analysis.
Innovation Solution
The use of a Convolutional Neural Network (CNN) architecture, combining one-dimensional or two-dimensional CNNs, Dilated CNNs, Gated Linear Units (GLUs), and Strided CNNs, is proposed as a mapping function to learn embedded vectors for sound source identification, stabilizing the learning process and improving sound source separation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a bidirectional long short-term memory (BLSTM) network is used as a mapping function to learn embedded vectors, then the model can capture temporal dependencies in spectrograms, but the learning process becomes unstable and prone to overlearning
Solution Approach 1:
The patent changes the fundamental parameters of the network architecture from recurrent connections (BLSTM) to convolutional operations (CNN). This parameter change transforms the network from one that processes sequences through hidden states to one that processes local patterns through filters, fundamentally altering the learning dynamics to achieve stability while maintaining effectiveness for spectrogram analysis
2Measurement precision
If a multi-layered RNN-based network is used, then the model can capture complex temporal patterns, but learning takes excessive time and overlearning occurs
Solution Approach 1:
The patent substitutes the recurrent neural network mechanism (which processes information sequentially through time steps) with a convolutional neural network mechanism (which processes information through parallel filter applications). This substitution replaces the sequential, memory-intensive RNN approach with a more efficient convolutional approach that achieves comparable or superior accuracy without the temporal computational burden
Solution Approach 2:
The patent segments the spectrogram processing into local receptive fields through convolutional filters, allowing parallel processing of different frequency-time regions. This segmentation enables the model to learn local patterns independently and combine them, achieving high accuracy without requiring deep temporal processing that would increase learning time
3Measurement precision
If consistent sound source labels are manually allocated to each spectrogram, then the training data quality improves, but the labor required increases significantly
Solution Approach 1:
The patent introduces an intermediary representation (embedded vectors) that automatically encodes sound source information without requiring manual labeling. Instead of directly labeling each spectrogram with consistent sound source identifiers, the system learns continuous vector representations that inherently capture sound source characteristics, enabling downstream clustering to identify consistent sources without manual intervention
Data Source
AI summary
It is possible to stably learn, in a short time, a model that can output embedded vectors for calculating a set of time frequency points at which the same sound source is dominant. Parameters of the neural network are learned based on a spectrogram of a sound source signal formed by a plurality of sound sources such that embedded vectors for time frequency points at which the same sound source is dominant are similar to embedded vectors for each of time frequency points output by a neural network, which is a CNN.