Phase-Aware Speech Enhancement Using Deep Recurrent Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech enhancement methods ignore phase information and assume a crude summation of magnitudes in the short-time Fourier transform domain, leading to suboptimal processing of noisy speech signals.
Innovation Solution
A deep recurrent neural network (DRNN) is used to predict a mask for the noisy speech signal, incorporating phase-sensitive processing and ASR features to enhance speech by learning network parameters that minimize the difference between clean and enhanced speech, allowing for phase-aware prediction of both magnitude and phase.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If phase information is ignored and magnitude summation is used in STFT domain, then processing complexity is reduced, but speech enhancement quality deteriorates
Solution Approach 1:
The patent changes the parameter representation from ignoring phase to explicitly modeling phase information. The phase prediction network outputs phase parameters that are combined with magnitude predictions to reconstruct the enhanced speech spectrum, transforming the crude magnitude-only approach into a phase-aware complex spectrum reconstruction process.
Solution Approach 2:
The patent introduces an intermediate phase prediction network that acts as a mediator between the input noisy speech and the final enhanced output. This network predicts phase information that serves as an intermediary component, which is then combined with magnitude predictions to form the complete enhanced speech representation.
2Manufacturing precision
If phase-aware prediction is implemented, then speech enhancement quality is improved, but computational requirements increase
Solution Approach 1:
The patent segments the speech enhancement task into two separate prediction networks: one for magnitude prediction and another for phase prediction. This segmentation allows each network to specialize in predicting specific parameters, improving overall accuracy while enabling efficient parallel computation and reducing the computational burden on a single network.
3Device complexity
If conventional magnitude-only prediction is used, then model simplicity is maintained, but speech recognition accuracy deteriorates
Solution Approach 1:
The patent adds another dimension to the prediction by incorporating phase information alongside magnitude. Instead of predicting only the magnitude spectrum, the system now predicts both magnitude and phase dimensions, creating a more complete representation of the speech signal that improves recognition accuracy without excessive complexity increase.
Data Source
AI summary
A method transforms a noisy audio signal to an enhanced audio signal, by first acquiring the noisy audio signal from an environment. The noisy audio signal is processed by an enhancement network having network parameters to jointly produce a magnitude mask and a phase estimate. Then, the magnitude mask and the phase estimate are used to obtain the enhanced audio signal.