Jahid05/bengali-hate-audio

Bengali Hate Speech Multimodal Features (BD-SHS Audio)

This dataset accompanies the paper:

"Phoneme-Level Hate Signals in Synthesized Bengali Speech: Complementary Signal Detection and Gate Collapse in Multimodal Hate Speech Fusion" Jahid [Last Name], Khulna University, 2025. Submitted to ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP).

---

Dataset Description

Pre-extracted multimodal features for Bengali hate speech detection, combining synthesized speech (Wav2Vec2) and text (BanglaBERT) representations derived from the BD-SHS dataset.

The BD-SHS text data is available at: https://github.com/naurosromim/hate-speech-dataset-for-Bengali-social-media

---

Files

File	Description	Shape / Size
metadata_reconstructed.csv	Maps audio filenames to labels and categories	27,696 rows
audio_features.npy	Wav2Vec2-base mean-pooled embeddings	(27696, 768)
text_features.npy	BanglaBERT CLS token embeddings	(27696, 768)
labels.npy	Binary hate speech labels (0=non-hate, 1=hate)	(27696,)
categories.npy	Topic category per sample	(27696,)
orig_idx.npy	Original BD-SHS dataset row indices	(27696,)
idx_train.npy	Training split indices (70%)	(19387,)
idx_val.npy	Validation split indices (15%)	(4154,)
idx_test.npy	Test split indices (15%)	(4155,)

---

Label Schema

Label	Class	Count	Percentage
0	Non-Hate	17,931	64.7%
1	Hate	9,765	35.3%

---

Category Distribution

Category	Total	Hate Rate
Crime	4,745	42.0%
Meme/TikTok	4,207	36.9%
Sports	4,392	36.3%
Entertainment	4,519	36.0%
Religion	4,357	36.0%
Politics	3,961	22.3%
Celebrity	3,662	20.2%

Note: 1,137 celebrity-category non-hate samples are absent from all splits due to API rate limiting during TTS synthesis. Celebrity-category results should be interpreted with this limitation in mind.

---

Feature Extraction Pipeline

Audio Features

• TTS Synthesis: Google Text-to-Speech (gTTS, lang='bn')
• Resampling: 22,050 Hz → 16,000 Hz (librosa)
• Model: facebook/wav2vec2-base
• Pooling: Mean pooling over time dimension of last hidden state
• Dimension: 768

Text Features

• Model: csebuetnlp/banglabert
• Extraction: CLS token from last hidden layer
• Max length: 128 tokens (truncated)
• Dimension: 768

Both models used in frozen inference mode — no fine-tuning performed.

---

Train/Val/Test Split

Stratified 70/15/15 split preserving label distribution. Split indices provided in idx_train.npy, idx_val.npy, idx_test.npy.

Split	Samples	Non-Hate	Hate
Train	19,387	12,552 (64.7%)	6,835 (35.3%)
Val	4,154	2,689 (64.7%)	1,465 (35.3%)
Test	4,155	2,690 (64.7%)	1,465 (35.3%)

---

Key Experimental Results

Model	F1-Macro	F1-Hate	AUC-ROC
Audio-only (Wav2Vec2)	0.6510	0.5649	0.7095
Text-only (BanglaBERT)	0.8247	0.7731	0.9103
Late Fusion	0.8225	0.7766	0.9137
Cross-Attention Fusion	0.8208	0.7746	0.9138
Gated Fusion	0.8222	0.7775	0.9112

Key finding: Gated fusion assigns only 3.46% weight to audio (gate collapse), despite audio carrying statistically significant complementary phoneme-level hate signal (binomial p<0.0001).

---

Usage

import numpy as np
from huggingface_hub import hf_hub_download

# Load features
audio_features = np.load(
    hf_hub_download('Jahid05/bengali-hate-audio',
                    'audio_features.npy', repo_type='dataset'))
text_features  = np.load(
    hf_hub_download('Jahid05/bengali-hate-audio',
                    'text_features.npy', repo_type='dataset'))
labels         = np.load(
    hf_hub_download('Jahid05/bengali-hate-audio',
                    'labels.npy', repo_type='dataset'))
idx_train      = np.load(
    hf_hub_download('Jahid05/bengali-hate-audio',
                    'idx_train.npy', repo_type='dataset'))
idx_test       = np.load(
    hf_hub_download('Jahid05/bengali-hate-audio',
                    'idx_test.npy', repo_type='dataset'))

print(f"Audio features: {audio_features.shape}")  # (27696, 768)
print(f"Text features : {text_features.shape}")   # (27696, 768)
print(f"Labels        : {labels.shape}")           # (27696,)

# Get train/test splits
X_audio_train = audio_features[idx_train]
X_text_train  = text_features[idx_train]
y_train       = labels[idx_train]

X_audio_test  = audio_features[idx_test]
X_text_test   = text_features[idx_test]
y_test        = labels[idx_test]

---

Reproducing Audio Synthesis

To re-synthesize the audio files from metadata:

import pandas as pd
from gtts import gTTS

meta = pd.read_csv('metadata_reconstructed.csv')

for _, row in meta.iterrows():
    tts = gTTS(text=row['text'], lang='bn', slow=False)
    tts.save(f"audio/{row['filename']}")

---

Citation

@article{jahid2025phoneme,
  title={Phoneme-Level Hate Signals in Synthesized Bengali Speech:
         Complementary Signal Detection and Gate Collapse in
         Multimodal Hate Speech Fusion},
  author={[Last Name], Jahid},
  journal={ACM Transactions on Asian and Low-Resource Language
           Information Processing},
  year={2025},
  note={Under review}
}

---

License

CC BY 4.0 — Free to use with attribution. Original BD-SHS dataset license applies to text content.