Real-Time Toxic Speech Detection Using Predictive Neural Networks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing audio signal processing systems face challenges in real-time detection and filtering of toxic speech in machine communication channels, leading to latency issues that are not suitable for applications like VoIP, where end-to-end latency requirements are stringent.

Innovation Solution

The system employs a combination of keyword spotting and predictive neural networks for real-time detection and modification of toxic words, using spectral masking to conceal the toxic parts of speech while allowing benign speech to pass through, thereby reducing latency by not waiting for the complete word to be recognized.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional speech recognition systems are used to detect and filter toxic speech, then toxic speech can be filtered out, but end-to-end latency increases significantly making the system unsuitable for real-time applications like VoIP

Engineering Contradiction:

Improvetoxic speech filtering effectivenessVSAvoidend-to-end latency

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The system performs preliminary detection of potentially toxic speech patterns before complete word recognition is finalized. By using keyword spotting and predictive neural networks to identify toxic patterns as they emerge, the system can intervene earlier in the speech stream, reducing the latency penalty while maintaining filtering effectiveness.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The speech processing pipeline is segmented into multiple stages: initial keyword spotting for rapid toxic pattern detection, predictive neural network analysis for confirmation, and selective filtering. This segmentation allows the system to process only suspicious segments in detail while letting benign speech pass through with minimal delay.

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If the system waits for complete word recognition before filtering, then accuracy improves, but latency increases making natural interaction difficult

Engineering Contradiction:

Improvetoxic speech detection accuracyVSAvoiddetection latency

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system applies partial recognition action by detecting keyword patterns and predictive features without waiting for full word completion. This partial action approach achieves sufficient detection accuracy for toxic speech while avoiding the latency of complete word recognition, allowing the system to act on partial evidence when toxicity is strongly indicated.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20220059071A1Sound modification of speech in audio signals over machine communication channels

Publication Date: 2022.02.24 INTEL CORP

AI summary

Apparatus, systems, articles of manufacture, and methods to modify sound of speech in an audio signal are disclosed. An example apparatus includes processor circuitry to execute instructions to: identify a first portion of a keyword in the speech of the audio signal during generation of the speech; determine a waveform to replace a second portion of the keyword; and transform the keyword into a different word by introducing the waveform into the audio signal.