Neural Voice Enhancement Using Low-Dimensional Speech Frames

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current voice enhancement methods distort speech features necessary for speech recognition and fail to enhance the quality, clarity, and intelligibility of unclear or degraded speech signals, particularly in noisy environments.

Innovation Solution

Utilizing a two-stage neural network approach, where a first neural network converts input audio data into low-dimensional representations and a second neural network generates target speech frames, preserving speech characteristics while suppressing noise, employing techniques like supervised and unsupervised learning to optimize the conversion process.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If speech separation processing is used to suppress all background sounds as noise, then noise reduction is improved, but speech features necessary for speech recognition are distorted

Engineering Contradiction:

Improvenoise reductionVSAvoidspeech features distortion

Core Design Contradiction:

Object-affected harmful factorsVSLoss of information

Solution Approach 1:

The system applies different processing characteristics to different frequency components and time segments of the audio signal. The neural network learns to selectively suppress background noise in certain frequency bands while preserving speech features in other bands, creating locally optimized noise reduction rather than uniform suppression across all frequencies.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The system dynamically adjusts processing parameters based on the input signal characteristics. The neural network continuously adapts its filtering behavior by changing parameters such as filter coefficients, gain values, and frequency band allocations in real-time, allowing optimal balance between noise reduction and speech preservation for each specific audio segment.

Inventive Principle:

Principle #35Parameter changes

2Object-affected harmful factors

If ratio mask technique is used to distinguish speech from background noise, then noise suppression is improved, but speech naturalness and clarity are degraded

Engineering Contradiction:

Improvenoise suppressionVSAvoidspeech clarity

Core Design Contradiction:

Object-affected harmful factorsVSManufacturing precision

Solution Approach 1:

The system extends the processing beyond simple frequency-domain ratio masking by incorporating temporal dimensions and multiple feature spaces. The neural network operates in a combined time-frequency domain and learns complex patterns across multiple dimensions, allowing more nuanced discrimination between speech and noise while preserving natural speech characteristics.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

3Manufacturing precision

If equalization is used to adjust frequency response for clarity, then speech clarity is improved, but speech features for recognition are distorted

Engineering Contradiction:

Improvespeech clarityVSAvoidspeech recognition features

Core Design Contradiction:

Manufacturing precisionVSLoss of information

Solution Approach 1:

The system uses dynamic, adaptive equalization rather than fixed frequency response adjustments. The neural network continuously adapts the equalization parameters based on the instantaneous characteristics of the input speech and noise, allowing the system to optimize clarity while preserving recognition features for each specific speech segment and environmental condition.

Inventive Principle:

Principle #15Dynamics

4Manufacturing precision

If noise suppression is applied to enhance speech quality, then perceived quality is improved, but speech intelligibility for unclear speech is reduced

Engineering Contradiction:

Improvespeech qualityVSAvoidspeech intelligibility

Core Design Contradiction:

Manufacturing precisionVSReliability

Solution Approach 1:

The system introduces an intermediate representation layer where the neural network learns to separate and reconstruct speech components. This intermediate processing stage allows the system to enhance quality through selective noise suppression while maintaining the integrity of speech features necessary for intelligibility, acting as a mediator between noise reduction and speech preservation.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS20260080886A1Methods for neural network-based voice enhancement and systems thereof

Publication Date: 2026.03.19 SANAS AI INC

AI summary

The disclosed technology relates to methods, voice enhancement systems, and non-transitory computer readable media for real-time voice enhancement. In some examples, input audio data including foreground speech content, non-content elements, and speech characteristics is fragmented into input speech frames. The input speech frames are converted to low-dimensional representations of the input speech frames. One or more of the fragmentation or the conversion is based on an application of a first trained neural network to the input audio data. The low-dimensional representations of the input speech frames omit one or more of the non-content elements. A second trained neural network is applied to the low-dimensional representations of the input speech frames to generate target speech frames. The target speech frames are combined to generate output audio data. The output audio data further includes one or more portions of the foreground speech content and one or more of the speech characteristics.