vitosa 1.0.0.post79

Vietnamese Audio-based Toxic Spans Detection Pipeline

ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances | INTERSPEECH'2025

---

Overview

ViToSA-Pipeline provides an end-to-end Python interface for:

• Transcribing Vietnamese audio into text using ViToSA PhoWhisper ASR.
• Detecting and masking toxic words with ViToSA PhoBERT TSD.
• Optionally processing full directories of .wav files.

It is built on top of the ViToSA 1.0 dataset, the first Vietnamese dataset for audio-based toxic spans detection, developed by researchers from the University of Information Technology (VNU-HCM).

---

About the ViToSA Dataset

Title: ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances
Version: 1.0 (August 2025)
Authors: Huy Ba Do, Vy Le-Phuong Huynh, and Luan Thanh Nguyen
Institution: University of Information Technology, VNU-HCM
Published: Interspeech 2025

If you use this dataset or models derived from it, please cite:

@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 in online platforms is a growing concern for user safety. While textual toxicity detection is well-studied, audio-based toxicity detection—especially for low-resource languages like Vietnamese—remains underexplored.

The ViToSA dataset introduces ~11,000 speech samples (≈25 hours) with human-annotated toxic spans, enabling ASR + Toxic Speech Detection (TSD) pipelines for Vietnamese. Experiments show that fine-tuning ASR models on ViToSA improves WER, and that text-based toxic span detection (TSD) models outperform baseline toxicity classifiers.

Dataset Summary

Split	# Examples	Notes
Train	8,641	toxic only
Validation	2,161	toxic only
Test	2,000	balanced 50/50 toxic / non-toxic

Each entry includes:

• file_name: audio file name
• audio: waveform (float32)
• transcript: human transcription
• toxicity: binary annotation

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

Modalities: Audio + Text
Language: Vietnamese
License: CC-BY-NC-ND 4.0 (non-commercial, no derivatives)

---

ViToSA Pipeline Features

✅ Automatic GPU / CPU selection
✅ Robust Hugging Face model handling
✅ Toxic word masking with "***"

---

Installation

pip install vitosa

---

Usage

1. Full Pipeline (ASR + TSD)

Transcribe and censor toxic words in an audio file:

from vitosa_pipeline import ViToSA

vitosa = ViToSA()
print(vitosa.pipeline("example.wav"))

2. Individual Components

2.1. ASR Only - Transcribe Audio

Transcribe an utterance with ViToSA-PhoWhisper:

from vitosa_pipeline import ViToSA

vitosa = ViToSA()
print(vitosa.asr("path_to_audio.wav"))

2.2. TSD Only - Detect Toxic Spans

Detect and mask toxic words with ViToSA-PhoBERT:

from vitosa_pipeline import ViToSA

vitosa = ViToSA()
transcript = "your_text_here"

# Return censored text with *** masking toxic words
print(vitosa.tsd(transcript))

# Return binary labels (0 = non-toxic, 1 = toxic word)
print(vitosa.tsd(transcript, return_labels=True))

---

Terms of Use

• Research and educational use only (non-commercial).
• Redistribution of raw data is prohibited.
• Cite the ViToSA paper if used in research.
• The dataset and models may contain explicit language. Use responsibly.

---

Contact

For more information: [e-mail]

---

MIT License © 2025 Luan Thanh Nguyen