Parametric Resynthesis for Speech Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech noise removal methods often inadvertently remove speech or fail to remove noise effectively, and speech synthesis systems struggle to predict prosody accurately from text inputs, leading to suboptimal results.
Innovation Solution
A parametric resynthesis method that predicts acoustic parameters from noisy speech signals using a neural network, which are then used to generate high-quality, noise-free speech through a vocoder, effectively combining speech enhancement and synthesis to produce more accurate prosody.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If traditional speech enhancement methods modify noisy speech directly, then noise removal is attempted, but speech quality deteriorates due to accidental removal of speech and incomplete noise removal
Solution Approach 1:
The patent introduces an acoustic model as an intermediary between the noisy speech and the final enhanced output. The model predicts clean speech acoustic features (spectrogram, pitch, energy) from the noisy input, acting as a mediator that separates speech from noise without directly modifying the original signal. This intermediary processing stage enables high-quality speech reconstruction while preserving natural prosody.
Solution Approach 2:
The patent replaces traditional signal processing methods (filtering, masking) with a neural network-based acoustic model. Instead of mechanical frequency-domain operations, the system uses deep learning to predict acoustic features, substituting conventional speech enhancement mechanics with intelligent prediction that better preserves speech quality.
2Productivity
If speech synthesis systems predict acoustic parameters from text, then speech generation is achieved, but prosody accuracy deteriorates due to difficulty in predicting timing, pitch, and intensity contours
Solution Approach 1:
The patent performs preliminary extraction of prosodic features (pitch contour, intensity contour, timing) from the noisy speech before synthesis. By capturing these features early in the processing pipeline and preserving them through the acoustic model prediction, the system establishes accurate prosody foundations before final speech generation, avoiding the need to predict them from scratch during synthesis.
Solution Approach 2:
The patent changes the approach from predicting prosody parameters from text to extracting and preserving them from the noisy speech signal itself. The acoustic model predicts clean speech parameters while maintaining the original prosodic characteristics, effectively changing the source of prosody information from textual prediction to signal-based preservation.
Data Source
AI summary
A method for Parametric resynthesis (PR) producing an audible signal. A degraded audio signal is received which includes a distorted target audio signal. A prediction model predicts parameters of the audible signal from the degraded signal. The prediction model was trained to minimize a loss function between the target audio signal and the predicted audible signal. The predicted parameters are provided to a waveform generator which synthesizes the audible signal.