The ViToSA 2.0 dataset contains LLM-augmented 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 2.0: A Multi-Task Approach Towards Robust Vietnamese Audio-Based Toxic Span Detection

This is the official repository for the ViToSA 2.0 dataset and model framework, introduced in the paper A Multi-Task Approach Towards Robust Vietnamese Audio-Based Toxic Span Detection, accepted at ICASSP 2026.
The dataset and multi-task framework were developed by researchers from the University of Information Technology, VNU-HCM.

Citation Information

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

@INPROCEEDINGS{11460540,
  author={Huynh, Vy Le-Phuong and Do, Huy Ba and Nguyen, Luan Thanh},
  booktitle={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  title={A Multi-Task Approach Towards Robust Vietnamese Audio-Based Toxic Span Detection},
  year={2026},
  volume={},
  number={},
  pages={18772-18776},
  keywords={Low earth orbit satellites;Feeds;Digital audio broadcasting;Broadcasting;Central Processing Unit;Electronic circuits;Radio access networks;Regional area networks;Protocols;HTTP;audio-based toxic span detection;multi-task learning;knowledge distillation;Vietnamese speech processing;content moderation},
  doi={10.1109/ICASSP55912.2026.11460540}
}

Abstract

Detecting toxic spans in spoken content is crucial for content moderation, but conventional two-stage pipelines that run Automatic Speech Recognition (ASR) followed by a Toxic Span Detection (TSD) model suffer from high latency and system complexity. To address these limitations, we propose a multi-task framework for Vietnamese audio-based toxic span detection that unifies ASR and TSD modules within a single model. Our approach utilizes a PhoWhisper encoder-decoder backbone and attaches a token classification head, enhanced with a BiLSTM and a Conditional Random Field (CRF) layer, directly to the decoder’s hidden states. To improve the model’s semantic understanding, we employ a knowledge distillation strategy to transfer rich textual knowledge from a pre-trained PhoBERT teacher model to our speech-based student model. Trained on a dataset augmented by a high-quality LLM-as-annotator-based pipeline, our model achieves a state-of-the-art F1-macro score of 0.9212. Furthermore, it significantly reduces inference latency by over 56% on GPU compared to the traditional two-stage approach, demonstrating a more efficient and robust solution for real-world deployment.

Data Structure

The dataset is distributed across two main Parquet files:

Split	File Name	File Size	# Examples
Train & Validation	`train_val.parquet`	~ 982 kB	[TO BE UPDATED]
Test	`test.parquet`	~ 91.7 kB	[TO BE UPDATED]
Total		~ 1.08 MB	[TO BE UPDATED]

Note: The train and validation sets are combined into a single train_val.parquet file.

Each entry consists of:
- file_name (string): The name of the audio file (e.g., audio1.wav).
- transcript (string): The speech transcription in Vietnamese.
- labels (string): A string representation of a list of binary labels indicating toxic spans at the token/word level (e.g., "[0, 1, 1, 0, ...]").

🔊 Example Usage: Load and Evaluate

Here’s a minimal example showing how to load the dataset from Hugging Face using the datasets library and parse the token-level labels:

import ast
import io
import numpy as np
import soundfile as sf
from datasets import load_dataset
from IPython.display import Audio, display

# 1. Load both dataset versions
ds_v1 = load_dataset("UIT-ViToSA/ViToSA-1.0")
ds_v2 = load_dataset("UIT-ViToSA/ViToSA-2.0")

# 2. Pick a sample from ViToSA 2.0 (test split)
v2_sample = ds_v2["test"][0]
target_filename = v2_sample["file_name"]
transcript = v2_sample["transcript"]
parsed_labels = ast.literal_eval(v2_sample["labels"])

print(f"Target Audio: {target_filename}")
print(f"Transcript:   {transcript}")
print(f"V2 Labels:    {parsed_labels}")

# 3. Lookup the corresponding audio in ViToSA 1.0 test split
df_v1_test = ds_v1["test"].to_pandas()
matched_row = df_v1_test[df_v1_test["file_name"] == target_filename].iloc[0]

# 4. Decode and listen to the audio
audio_bytes = matched_row["audio"]["bytes"]
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)

display(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.

License

This dataset is released under the Apache 2.0 License.

Repository

GitHub: https://github.com/vitosa-research/ViToSA-Dataset

Contact

Please feel free to contact us by email [email protected] if you have any further information!

Downloads last month: 43

Total file size:

1.08 MB

Collection including UIT-ViToSA/ViToSA-2.0

ViToSA 2.0

Collection

2 items • Updated 19 days ago