AcousticToxicityDetector<T>

Models & Types · AiDotNet.Safety.Audio

Detects toxic/aggressive speech patterns directly from acoustic features without transcription.

Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → AudioSafetyModuleBase<T> → AcousticToxicityDetector<T>

For Beginners

You can often tell someone is angry or threatening just from the tone of their voice — even if you don’t understand the language. This module does the same: it analyzes voice patterns like loudness changes, pitch, and speaking rate to detect aggressive or hostile speech.

How It Works

Analyzes prosodic and spectral features that correlate with aggression, shouting, and hostile intent. Aggressive speech has characteristic acoustic signatures: elevated pitch, high energy variance, rapid pitch changes, and spectral centroid shifts. This approach works regardless of language and catches toxicity even when words are unintelligible.

References:

  • Taxonomy of speech generator harms including swatting attacks (2024, arxiv:2402.01708)
  • Acoustic emotion recognition using prosodic and spectral features (2023)
  • Paralinguistic analysis for aggression detection (INTERSPEECH 2024)

Constructors

Constructor Summary
AcousticToxicityDetector(double threshold, int sampleRate) Initializes a new acoustic toxicity detector.

Properties

Property Summary
string ModuleName { get; }

Methods

Method Summary
IReadOnlyList<SafetyFinding> EvaluateAudio(Vector<T> audioSamples, int sampleRate)