Neural Network Signal Generation for Speech Verification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional signal processing techniques using neural networks face challenges in verifying the correctness of speech enhancement operations, as they lack direct information about frequency, which is essential for generating accurate time-series acoustic signals.

Innovation Solution

A signal generation device that extracts frequency information from the weights of a neural network's hidden layers, using this information along with amplitude and phase to generate time-series signals, allowing for the verification of speech enhancement operations by converting these signals into audible outputs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If neural network is used for speech enhancement, then speech processing capability is improved, but verification of processing correctness becomes difficult

Engineering Contradiction:

Improvespeech processing capabilityVSAvoidverification of processing correctness

Core Design Contradiction:

ProductivityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent introduces an intermediary signal generation device that acts as a mediator between the neural network and the verification process. This device extracts frequency information from the neural network's hidden layers and generates time-series signals that can be aurally verified, thus bridging the gap between complex neural network processing and human-verifiable outputs

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces the abstract neural network processing with a more tangible acoustic signal generation system. By converting the hidden layer outputs into audible time-series signals, it substitutes the difficult-to-verify neural network computations with perceptible acoustic outputs that can be directly evaluated by humans

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Adaptability or versatility

If frequency information is extracted from neural network weights, then time-series signal generation is enabled, but device complexity increases

Engineering Contradiction:

Improvetime-series signal generation capabilityVSAvoidsignal generation device structure

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent extracts frequency information directly from the neural network's hidden layer weights, separating this critical frequency data from the rest of the complex neural network structure. This extraction enables time-series signal generation while isolating the complexity management to a dedicated frequency extraction module

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The signal generation device is designed to handle multiple functions: it processes amplitude and phase information from the neural network, extracts frequency data from weights, and generates time-series signals. This multi-functionality consolidates what would otherwise require separate systems into a single versatile device

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS11282505B2Acoustic signal processing with neural network using amplitude, phase, and frequency

Publication Date: 2022.03.22 KK TOSHIBA

AI summary

According to one embodiment, a signal generation device includes one or more processors. The processors convert an acoustic signal and output amplitude and phase at a plurality of frequencies. The processors, for each of a plurality of nodes of a hidden layer included in a neural network that treats the amplitude and the phase as input, obtain frequency based on a plurality of weights used in arithmetic operation of the node. The processors generate an acoustic signal based on the plurality of obtained frequencies and based on amplitude and phase corresponding to each of the plurality of nodes.