💬 NLP

Few-Shot Contrastive Adaptation for Audio Abuse Detection in Low-Resource Indic Languages

This paper demonstrates that few-shot supervised contrastive adaptation of CLAP models enables effective cross-lingual abusive speech detection in low-resource Indic languages directly from audio, achieving competitive performance with fully supervised systems while highlighting that adaptation benefits vary non-monotonically across languages.

Original authors: Aditya Narayan Sankaran, Reza Farahbakhsh, Noel Crespi

Published 2026-04-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aditya Narayan Sankaran, Reza Farahbakhsh, Noel Crespi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a moderator for a massive, global town square where people are shouting, singing, and arguing in dozens of different languages. Your job is to spot the bullies and the abusers.

In the past, the only way to do this was to hire a translator for every single language. You'd listen to the shout, write it down in English (transcription), and then ask a text-expert to decide if it was mean. But this system has two big problems:

The translator makes mistakes: If the speaker is shouting, mumbling, or mixing languages (like speaking Hindi and English at the same time), the translator gets confused. If the translation is wrong, the bully gets a free pass.
You lose the "tone": A sentence like "You're a genius" can be a compliment or a sarcastic insult. The words are the same, but the tone (the pitch, the anger, the speed) tells the real story. Translators throw away the tone.

The New Idea: The "Universal Ear"

This paper proposes a smarter way. Instead of translating first, they use an AI called CLAP (Contrastive Language-Audio Pre-training). Think of CLAP as a super-ear that has listened to millions of hours of audio and read millions of books. It doesn't just hear words; it understands the vibe and the meaning of sounds directly.

The researchers wanted to see if this "super-ear" could spot abuse in Indic languages (languages of India like Hindi, Tamil, Bengali, etc.) without needing a huge library of labeled examples for every single one.

The Experiment: The "Few-Shot" Challenge

Usually, to teach an AI to spot a bully, you need thousands of examples of "bad" and "good" speech for that specific language. But in low-resource languages, you might only have a handful of examples. This is called the "Few-Shot" problem.

The researchers asked: Can we teach this super-ear to spot abuse in a new language just by showing it a tiny handful of examples (like 5 or 25)?

They tried two main strategies:

The "Light Touch" (Projection-only): They kept the super-ear's brain frozen and just added a tiny, adjustable "adapter" on top of it. It's like giving the AI a pair of glasses specifically tuned for that language, without changing how it sees the world.
The "Heavy Lifting" (Fine-tuning): They let the AI re-learn parts of its brain to fit the new language. This is like trying to rewire the AI's entire nervous system.

What They Found (The Results)

1. The "Light Touch" Wins
Surprisingly, the simple "Light Touch" method worked just as well, and sometimes even better, than the heavy re-wiring.

Analogy: Imagine you are an expert chef (the pre-trained AI). You don't need to learn how to cook a whole new cuisine from scratch (fine-tuning). You just need to swap out the salt shaker for a specific spice blend (the projection head) to make the dish perfect.
The Result: For languages like Punjabi, the "Light Touch" method actually did better than systems trained on massive amounts of data!

2. It's Not One-Size-Fits-All
The magic of this AI depends on the language.

For some languages, showing the AI just 5 examples was enough to make it a pro.
For others, showing it 25 examples helped, but showing it zero examples (just guessing based on its general knowledge) was actually the best!
Metaphor: It's like learning to drive. In a quiet neighborhood (some languages), you only need a few minutes of instruction. In a chaotic city (other languages), you might need a full driving school. But for some specific roads, you might actually be safer just driving on instinct (zero-shot) than overthinking it.

3. The "Leave-One-Out" Test
They also tested if the AI could spot abuse in a language it had never seen before, by training it on all the other languages and then testing it on the new one.

The Result: The AI was surprisingly good at this. It learned the "universal signs of anger" (like a rising pitch or a sharp rhythm) that exist across many languages. However, it wasn't perfect. It still needed a little bit of help from the specific language to be truly accurate.

The Big Takeaway

This paper proves that we don't need to build a massive, expensive factory for every single language to detect online abuse.

We can use a universal "super-ear" that understands the feeling of abuse across languages. By giving it just a tiny nudge (a few examples) and a simple adapter, it can become a highly effective moderator for low-resource languages.

In short: Instead of translating the shout and hoping for the best, we can now listen to the tone of the shout directly, even if we've only heard that specific language a few times before. It's a faster, cheaper, and more human way to keep our digital town squares safe.

1. Problem Statement

The paper addresses the critical challenge of detecting abusive speech in low-resource, multilingual settings, specifically focusing on ten Indic languages.

Limitations of Current Approaches: Most existing systems rely on an Automatic Speech Recognition (ASR) pipeline followed by text-based hate speech classification. This approach suffers from:
- Transcription Errors: ASR models for low-resource and code-mixed languages often fail, obscuring abusive intent.
- Loss of Prosody: Crucial paralinguistic cues (pitch, stress, emotional intensity) that define abuse are discarded when audio is converted to text.
- Error Propagation: Errors in the ASR stage compound in the downstream classification stage.
The Gap: There is a lack of robust, audio-native detection systems that can operate directly on speech signals without transcription, particularly in low-resource scenarios where large annotated datasets are unavailable.

2. Methodology

The authors propose a framework based on Contrastive Language-Audio Pre-training (CLAP) to detect abuse directly from audio using few-shot learning.

Core Architecture

Base Model: They utilize the pre-trained CLAP model, which learns aligned audio and text embeddings in a shared semantic space using a symmetric InfoNCE loss.
Input Processing: Audio is resampled and padded/truncated to a fixed duration before being encoded into 512-dimensional embeddings.

Experimental Settings

The study evaluates the model under three distinct transfer scenarios:

Zero-Shot Prompting: Classifying audio by computing cosine similarity between the audio embedding and text prompts (e.g., "This audio contains hate speech" vs. "This audio does not contain hate speech").
Few-Shot Supervised Contrastive Adaptation: Using a small support set ( $k$ $k$ examples per class per language) to refine the embedding space. Two adaptation strategies are tested:
- Projection-Only: The CLAP encoder is frozen; only a lightweight projection head is trained using supervised contrastive loss.
- Projection + Fine-Tuning: The projection head and the final blocks of the audio encoder are fine-tuned jointly.
Downstream Classification: The adapted embeddings are fed into two classifiers: a Support Vector Machine (SVM) and a Multi-Layer Perceptron (ANN-MLP). The best-performing classifier is selected per language.

Evaluation Protocols

Dataset: The ADIMA dataset, containing 11,775 audio clips across 10 Indic languages (Bengali, Bhojpuri, Gujarati, Haryanvi, Hindi, Kannada, Malayalam, Odia, Punjabi, Tamil).
Generalization Settings:
- Monolingual: Train and test on the same language.
- Cross-Lingual: Train on all languages, test on specific targets.
- Leave-One-Language-Out (LOLO): Train on all languages except the target, testing the model's ability to generalize without any target-language supervision.

3. Key Contributions

Audio-Native Detection: Demonstrates that CLAP-based representations can effectively detect abusive speech directly from audio, bypassing the error-prone ASR pipeline.
Few-Shot Efficiency: Shows that projection-only adaptation (updating only a small projection head) achieves competitive performance with fully supervised systems trained on complete datasets, using as few as 5–25 labeled examples per class.
Cross-Lingual Analysis: Provides a comprehensive analysis of how well abusive speech cues transfer across diverse Indic languages, revealing that transfer is effective but not uniform.
LOLO Insights: Introduces a rigorous Leave-One-Language-Out analysis to quantify the dependency on target-language supervision, finding that while cross-lingual training is generally superior, the model captures significant language-agnostic abuse cues.

4. Key Results

Performance vs. Full Supervision:
- The CLAP-based few-shot approach narrows the gap to fully supervised systems significantly.
- Punjabi: Projection-only adaptation achieved 83.65% Macro-F1 (at 5-shot), slightly outperforming the fully supervised ADIMA benchmark of 83.40%.
- Other Languages: Strong results were observed for Haryanvi (80.23%), Odia (79.67%), and Malayalam (78.18%) with limited shots.
- For languages with larger gaps (e.g., Bengali, Hindi), the few-shot results remained within 2–4 Macro-F1 points of full supervision, which is a significant achievement given the data constraints.
Adaptation Strategy:
- Projection-only adaptation generally proved more stable and effective than full encoder fine-tuning, suggesting that the pre-trained CLAP space already contains rich task-specific information.
- The benefits of few-shot adaptation are non-monotonic; increasing the shot size does not always improve performance (e.g., some languages performed best at 0-shot or 25-shot, not 50-shot).
LOLO vs. Cross-Lingual:
- LOLO performance remained competitive with standard cross-lingual training (e.g., mean Macro-F1 of 76.99 vs. 77.18 for projection-only), indicating the model learns language-agnostic acoustic cues of abuse.
- However, cross-lingual training retained a small but persistent advantage, suggesting that transfer is incomplete and some languages still benefit from direct exposure.

5. Significance and Implications

Scalability for Low-Resource Regions: This work provides a practical pathway for deploying content moderation systems in linguistically diverse regions (like South Asia) where collecting large annotated datasets is infeasible.
Paralinguistic Preservation: By operating directly on audio, the system preserves emotional and prosodic cues essential for detecting impulsive or emotionally charged abuse that text-based models miss.
Efficiency: The "projection-only" strategy offers a computationally efficient compromise, requiring minimal parameter updates while leveraging massive pre-trained knowledge.
Future Directions: The paper highlights that while contrastive models are promising, transfer is language-dependent. Future work should focus on better prompt engineering, parameter-efficient adaptation, and deeper linguistic analysis of transfer behaviors.

Conclusion: The study validates that CLAP-based audio-text models are a robust foundation for cross-lingual audio abuse detection. It demonstrates that with minimal supervision (few-shot), these models can achieve performance comparable to fully supervised systems, offering a scalable solution for moderating harmful speech in low-resource Indic languages.