Version 1.0 — August 2025
The ViToSA dataset contains Vietnamese speech recordings and transcripts with annotated toxic spans, created by researchers at the University of Information Technology, VNU-HCM.
⚠️ Disclaimer: This dataset includes explicit and offensive language. Use responsibly for research on content moderation and language modeling.

Terms of Use

Use of the Dataset is permitted only for non-commercial research and educational purposes.
Redistribution of raw data (audio, transcripts, or annotations) is not allowed.
You must properly cite the original paper.
Do not attempt to identify individuals or use the Dataset to harass, discriminate, or harm.
The Dataset is provided “as is” without warranty.
Access may be revoked if these terms are violated, and you must delete all copies.
By requesting access, you confirm that you have read and agreed to these terms.

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

This is the official repository for the ViToSA 1.0 dataset, introduced in the paper ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances, accepted at Interspeech 2025.
The dataset was developed by researchers from the University of Information Technology, VNU-HCM.

Citation Information

If you use this dataset, please cite the following paper:

    @inproceedings{do25b_interspeech,
      title     = {{ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances}},
      author    = {Huy Ba Do and Vy Le-Phuong Huynh and Luan Thanh Nguyen},
      year      = {2025},
      booktitle = {{Interspeech 2025}},
      pages     = {4013--4017},
      doi       = {10.21437/Interspeech.2025-1958},
      issn      = {2958-1796},
    }

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.

Data Distribution

Split	# Examples (Toxic & None)
Train	8,641 – 0
Validation	2,161 – 0
Test	1,000 – 1,000
Total	12,802

Note: Only the test split includes both toxic and non-toxic utterances (balanced 50/50); The train and validation sets currently contain only toxic examples; The train and validation sets are split with an 8:2 ratio.

Each entry consists of:
- file_name: audio file name
- audio: audio waveform (float32)
- transcript: speech transcription in Vietnamese
- toxicity: binary annotation (toxic or non-toxic)

For the test set, we also include its annotated toxic spans.

The dataset provides a foundation for developing audio-based toxic spans detection systems in Vietnamese, complementing existing text-based datasets.

🔊 Example Usage: Load and Play a Sample

Here’s a minimal example showing how to load and play a sample audio file from the ViToSA dataset:

import io
import soundfile as sf
import numpy as np
from IPython.display import Audio

# Load dataset and access first test sample
audio_bytes = ds["test"].to_pandas().iloc[0]["audio"]["bytes"]

# Decode and preprocess audio
data, sr = sf.read(io.BytesIO(audio_bytes))
if data.ndim > 1:
    data = data.mean(axis=1)
data = np.clip(data, -1, 1).astype(np.float32)

# Play audio
Audio(data, rate=sr)

Dataset Access Agreement

⚠️ Disclaimer: This dataset contains explicit and offensive language.
Use responsibly for research on content moderation and language modeling. Access is restricted to non-commercial research and educational use under the following terms:

Redistribution of raw data is not allowed
Proper citation is required (see above)
Any violation of the terms may result in access revocation Full agreement text is provided in the gated access form.