PaperlayerView exact revision on arXiv

Source-linked AI summary

ToxicTone: A Mandarin Audio Dataset Annotated for Toxicity and Toxic Utterance Tonality

Yu-Xiang Luo, Yi-Cheng Lin, Ming-To Chuang, Jia-Hung Chen, I-Ning Tsai, Pei Xing Kiew, Yueh-Hsuan Huang, Chien-Feng Liu, Yu-Chen Chen, Bo-Han Feng, Wenze Ren, Hung-yi Lee

arXiv:2505.15773v1eess.AScs.CLcs.SD

TL;DR

Spoken Mandarin toxicity detection lacks large labeled datasets that capture spoken-language and culture-specific features. ToxicTone introduces a public dataset with detailed toxicity labels and combines acoustic, emotional, and semantic information; the authors report better detection performance than text-only approaches.

  • Problem

    Mandarin toxic speech research has scarce labeled datasets and is largely text-focused, despite spoken communication and culturally specific expressions presenting distinct challenges.

    Source: §1 · p. 1

  • Method

    ToxicTone provides a Mandarin toxic speech dataset with labels for toxicity form and source, while the detection approach combines speech, emotion, and semantic encoders.

    Source: §1 · p. 1

  • Results

    Combining speech, emotion, and semantic encoders achieves the best results and outperforms text-only approaches in toxic speech detection.

    Source: §1 · p. 1

  • Takeaways & Limitations

    The reported findings highlight the importance of prosodic and emotional information and support speech-focused methods for detecting toxicity.

    Source: §1 · p. 1

    The experiments reflect majority opinion, which may not fully capture toxicity perceptions shaped by personal experience, culture, and context.

    Source: §5 · p. 4

Abstract

from arXiv · show

Despite extensive research on toxic speech detection in text, a critical gap remains in handling spoken Mandarin audio. The lack of annotated datasets that capture the unique prosodic cues and culturally specific expressions in Mandarin leaves spoken toxicity underexplored. To address this, we introduce ToxicTone -- the largest public dataset of its kind -- featuring detailed annotations that distinguish both forms of toxicity (e.g., profanity, bullying) and sources of toxicity (e.g., anger, sarcasm, dismissiveness). Our data, sourced from diverse real-world audio and organized into 13 topical categories, mirrors authentic communication scenarios. We also propose a multimodal detection framework that integrates acoustic, linguistic, and emotional features using state-of-the-art speech and emotion encoders. Extensive experiments show our approach outperforms text-only and baseline models, underscoring the essential role of speech-specific cues in revealing hidden toxic expressions.

1. Introduction

The paper addresses gaps in Mandarin toxic speech detection by introducing ToxicTone, a public dataset with detailed toxicity labels and prosodic cues. It also reports better detection performance from combining speech, emotion, and semantic encoders than from text-only approaches.

1. Introduction

  • Mandarin spoken toxicity detection remains underdeveloped amid scarce labeled datasets, text-focused research, and culturally specific colloquialisms and code-switching.Existing Chinese datasets may lack labels that distinguish toxic speech from general profanity or mark hidden insults.Source: §1 · p. 1
  • Combining speech, emotion, and semantic encoders achieves the best results, outperforming text-only approaches in toxic speech detection.The authors argue this highlights the need to include prosodic and emotional information in speech.Source: §1 · p. 1
  • ToxicTone is presented as the largest public Mandarin Chinese toxic speech dataset, with labels for both the form and source of toxicity.Its annotations include prosodic features such as intonation, emphasis, and rhythm that may signal harmful intent not visible in text.Source: §1 · p. 1
  • The paper aims to support safer online spaces and advance spoken-language toxicity research through its data and modeling contributions.Source: §1 · p. 1

2. Related work

Prior Chinese toxicity datasets primarily address text, while existing speech datasets have limits in language coverage, sample size, source, or toxicity scope.

2. Related work

  • Chinese datasets such as COLA, TOCP, COLD, and ToxiCN classify toxicity in text, with differing sample sizes and label schemes.COLA has 18.7k comments, TOCP 16k, COLD 37k, and ToxiCN 12k hierarchical annotations.Source: §2 · p. 1
  • Text-based datasets miss prosodic cues such as intonation, pitch, and stress, as well as spontaneous language, emotion, and interruption in real-world interactions.Source: §2 · pp. 1–2
  • Existing speech datasets are constrained by English-only data, small podcast-based multilingual samples, or a focus on profanity in Indic languages.Detoxy is English-only; Mutox has 2,000 samples per language, primarily from podcasts; ADIMA detects profanity.Source: §2 · p. 2

3. Dataset collection

ToxicTone defines toxicity by both its form and source, then builds the dataset from web-crawled speech through filtering and human annotation. The resulting dataset is the largest publicly available toxic speech detection dataset, with imbalanced labels across 13 topical categories.

3.1. Definition of toxicity

  • Toxicity is classified by form, such as profanity, hate speech, pornography, bullying, sarcasm, or other toxic speech, and by source, such as specific words, angry or violent tone, dismissiveness, sarcasm, or threats.Form describes the manifestation of harmful language; source describes the origin of harmful intent or tone.Source: §3.1 · p. 2

3.2. Preprocessing

  • Web-crawled recordings were diarized, transcribed, split into 2–10-second clips, then filtered by transcript toxicity score, retaining scores above 0.75.Filtering reduced 770k clips to 52k, with a rule-based word list adding 600 samples potentially exhibiting pornographic toxicity.Source: §3.2 · p. 2

3.3. Human Annotation

  • Eleven native Chinese annotators labeled samples for one or more toxicity forms and corresponding sources, or marked them non-toxic; tied toxic/non-toxic votes received another review.The annotation took about 1.5 months and 900 collective hours.Source: §3.3 · p. 2
  • ToxicTone is the largest publicly available toxic speech detection dataset and is divided into train, development, and test sets.Its form and source labels are imbalanced: bullying speech appears nearly 5,800 times versus fewer than 400 pornographic-language samples; specific-word labels appear almost 8,000 times versus about 560 threatening-tone samples.Source: §3.3 · p. 2

3.4. Category distribution

  • The dataset spans 13 topical categories, assigned using GPT-4o mini on transcripts from the first 30 minutes of each recording before segmentation.Society & News and Entertainment are the largest categories, with 14,247 and 13,684 clips, respectively; Gaming has 8,326 clips, including 4,133 toxic clips.Source: §3.4 · pp. 2–3

4. Experiments

The experiments evaluate binary toxicity detection and toxicity-source classification using speech, text, and emotion representations, alone and in combination. The multimodal configuration achieves the strongest toxicity-detection results, while source-classification performance varies by category.

4.1. Experiment Type

  • The experiments define toxicity detection as binary classification and source classification as identifying the origin of toxicity, including aggressive wording, sarcastic tone, or threatening intent.Figure 2 reports total and toxic clip counts by category.Source: §4.1 · p. 3, Figure 2 · p. 3

4.2. Experiment Setup

  • The setup compares MuTox, ETOX, and COLDETECTOR with SONAR text and speech encoders, XLS-R 1B, Emotion2Vec+, and concatenated feature combinations.Downstream tasks use three-layer linear classifiers; source categories are classified with separate one-vs-all models.Source: §4.2 · pp. 3–4

4.4. Toxicity source detection

  • ST+E+X leads on Angry/Violent and Sarcastic/Satirical, but semantic encoder ST performs best on threat F1 despite uniformly high threat accuracy.Threat-labeled samples are considerably fewer, which the authors identify as a reason for the lower F1.Source: Figure 3 · p. 4, §4.4 · p. 4

4.3. Toxicity detection

  • X+ST+E attains the best toxicity-detection results, while X is the strongest single encoder and ST+E improves on individual encoders.ST+E yields an F1 of 62.47% and recall of 70.09%; X reaches an F1 of 60.89%.Source: Table 3 · p. 4, §4.3 · p. 4

5. Limitation and future work

The dataset preserves real-world class imbalances, including higher toxic-speech prevalence in gaming content, and some categories remain underrepresented. The authors also note that majority-opinion labels may not capture individual differences in toxicity perception.

  • The dataset reflects natural class imbalances, including a higher prevalence of toxic speech in gaming content.Source: §5 · p. 4
  • The authors identify underrepresented categories as a coverage boundary and propose expanding them to improve generalization.Source: §5 · p. 4
  • Majority-opinion annotations may not capture perceptions shaped by personal experience, culture, and context.Source: §5 · p. 4

6. Conclusion

ToxicTone establishes a large-scale benchmark for Mandarin spoken toxicity, and experiments find multimodal approaches outperform text-only models. The conclusion also identifies gaming content as having the highest toxicity prevalence and points to future dataset and model development.

6. Conclusion

  • ToxicTone provides a large-scale benchmark for Mandarin spoken toxicity, and multimodal approaches significantly outperform text-only models.The dataset includes prosodic cues and detailed toxicity labels; the experiments combine acoustic, emotional, and linguistic features.Source: §6 · p. 4
  • Gaming content exhibits the highest prevalence of toxicity among the domains analyzed.Source: §6 · p. 4
  • The work lays a foundation for multimodal toxicity detection, Chinese dialect dataset expansion, and personalized toxicity-perception models for online speech moderation.Source: §6 · p. 5
Loading 2505.15773v1…