Hate Speech Classification in Audio

May 24, 2024

• Python
• PyTorch
• Transformers
• Wav2Vec2
• AST
• DistilBERT
• Coqui-TTS

• AST F1: 0.69
• Wav2Vec2 F1: 0.68
• DistilBERT F1: 0.77 (text baseline for comparison)
• Grade: 5.25 / 6

TL;DR

• Compared waveform-based (Wav2Vec2), spectrogram-based (AST), and text-based (DistilBERT) models for hate speech detection.
• DistilBERT on raw text achieved the best performance (F1 = 0.77).
• Audio models (AST, Wav2Vec2) performed slightly worse (F1 ≈ 0.69 / 0.68) but showed promise given only synthesized TTS data.
• Highlighted the potential of audio-based hate detection, especially for future datasets with real human speech.

At a glance

• Role: Deep learning researcher (team of 3)
• Timeline: Apr–May 2024 (6 weeks)
• Context: EPFL Deep Learning mini-project
• Users/Stakeholders: Moderation systems, researchers studying speech-based detection
• Scope: Dataset synthesis with TTS, AST and Wav2Vec2 fine-tuning, Hugging Face hosting

Problem

Detecting hate speech from speech audio is underexplored compared to text classification. Challenges:

1. Lack of labeled audio hate speech datasets.
2. Balancing waveform vs spectrogram representations for model input.
3. Establishing a fair comparison with text-based baselines.

The project asked: Can audio-only models reliably detect hate speech, and how do they compare to established text classifiers?

Solution overview

We synthesized a speech dataset from social media comments using Coqui-TTS and trained three models:

1. Wav2Vec2 — raw waveform classification.
2. AST (Audio Spectrogram Transformer) — spectrogram-as-image classification.
3. DistilBERT — text baseline for comparison.

All models were fine-tuned with Hugging Face’s Trainer API and evaluated on accuracy, precision, recall, and F1.

Architecture

• Dataset pipeline: Social media comments → preprocessing (remove hashtags, emojis, mentions) → balanced hate vs non-hate labels → TTS synthesis into audio samples.
• Models:
  • Wav2Vec2 (raw waveform encoder + classifier).
  • AST (spectrograms treated as images, transformer-based).
  • DistilBERT (text sequence classification baseline).
• Evaluation: Binary classification across validation/test splits.

Data

• Source: UC Berkeley Measuring Hate Speech dataset (~39k comments).
• Labels: Continuous hate score → binarized into hate vs non-hate.
• Synthesis: Audio generated via Coqui-TTS (Jenny model), producing clear speech samples.
• Splits: Train/validation/test; validation used for hyperparameter tuning.

Method

Wav2Vec2 (waveform-based)

• Model: facebook/Wav2Vec2-base.
• Fine-tuned for binary classification.
• Pros: direct speech representations.
• Cons: limited by training compute; ~94M parameters.

AST (spectrogram-based)

• Model: MIT/ast-finetuned-audioset.
• Input: spectrograms treated as 2D images.
• Pros: efficient (~86M params), converged faster.
• Cons: requires preprocessing pipeline for spectrograms.

DistilBERT (text baseline)

• Distilled transformer (~66M params).
• Fine-tuned on raw text samples from the dataset.
• Served as upper bound for comparison.

Experiments & Results

Benchmarks

Evaluation protocol.

• Metric: Accuracy, Precision, Recall, F1.
• Hyperparameters tuned per model (learning rate, batch size, gradient accumulation).
• Best checkpoint selected by F1 on validation set.

Analysis

• DistilBERT unsurprisingly outperformed audio-only models.
• AST slightly ahead of Wav2Vec2 — likely due to its smaller size and faster convergence.
• Audio models still showed promise, with only ~0.08 drop in F1 vs text, despite relying on synthetic TTS audio.
• Limitation: lack of real human voices; synthetic audio reduces robustness.

Impact

• Demonstrated that audio-based hate speech detection is viable and approaches text-based baselines.
• Highlighted potential for speech moderation tools where only audio streams are available (e.g., podcast, voice chat, music).
• Exposed limitations of synthetic data, motivating the need for real annotated speech datasets.
• Released all resources openly on Hugging Face for reproducibility:
  • AST model
  • Wav2Vec2 model
  • DistilBERT model
  • Synthesized dataset

What I learned

• Designing fair benchmarks requires keeping datasets constant across modalities.
• Training resource constraints strongly shape model choice (AST vs Wav2Vec2).
• TTS-based data generation can bootstrap research, but realism matters for generalization.
• Comparative analysis sharpened my understanding of representation learning across modalities.

Future Work

• Collect or integrate real human-voice hate speech datasets.
• Explore multi-modal models combining audio, text, and video.
• Investigate self-supervised speech encoders fine-tuned specifically for toxicity.
• Improve robustness with background noise, varied voices, and real-world conditions.

References

• Kennedy et al. (2020): Measuring Hate Speech dataset.
• Gong et al. (2021): Audio Spectrogram Transformer (AST).
• Baevski et al. (2020): Wav2Vec2.0.
• Sanh et al. (2019): DistilBERT.
• Coqui-TTS documentation.