ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances

2025-05-31

AI Summary

Audio-based toxic content detection remains unaddressed for low-resource languages like Vietnamese. Method: We propose the first end-to-end framework for fine-grained toxic speech segment detection. To support this, we introduce ViToSA—the first multimodal Vietnamese toxic speech dataset—comprising 11,000 annotated audio clips (25 hours), each labeled with aligned transcriptions and toxic span boundaries. Our pipeline jointly leverages ASR fine-tuning (Whisper/Wav2Vec 2.0) and sequence labeling–based toxic segment detection (BERT-CRF/Span-based TSD). Contribution/Results: Experiments show substantial WER reduction for ASR on ViToSA and an 8.2%-point F1 improvement for TSD over baselines. This work establishes the first benchmark for toxic speech detection in low-resource languages and provides a reproducible paradigm for audio content moderation.

Technology Category

• Natural Language Processing: Speech
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Language and Vision

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Toxic speech on online platforms is a growing concern, impacting user experience and online safety. While text-based toxicity detection is well-studied, audio-based approaches remain underexplored, especially for low-resource languages like Vietnamese. This paper introduces ViToSA (Vietnamese Toxic Spans Audio), the first dataset for toxic spans detection in Vietnamese speech, comprising 11,000 audio samples (25 hours) with accurate human-annotated transcripts. We propose a pipeline that combines ASR and toxic spans detection for fine-grained identification of toxic content. Our experiments show that fine-tuning ASR models on ViToSA significantly reduces WER when transcribing toxic speech, while the text-based toxic spans detection (TSD) models outperform existing baselines. These findings establish a novel benchmark for Vietnamese audio-based toxic spans detection, paving the way for future research in speech content moderation.

Problem

• Detecting toxic spans in Vietnamese speech audio
• Addressing lack of audio-based toxicity datasets for Vietnamese
• Improving ASR and TSD models for toxic speech content

Innovation

• First Vietnamese toxic speech dataset ViToSA
• Pipeline combining ASR and toxic spans detection
• Fine-tuning ASR reduces toxic speech WER