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

Aditya Narayan Sankaran; Reza Farahbakhsh; Noel Crespi

cs.SDApr 10, 2026

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

Authors: Aditya Narayan Sankaran, Reza Farahbakhsh, Noel Crespi

Organizations: SAMOVAR, Télécom SudParis Institut Polytechnique de Paris 91120 Palaiseau, France

Abstract

Abusive and hateful speech is increasingly spoken rather than written, surfacing in voice notes, calls, and short-form videos. Most detection systems still transcribe speech to text before classifying it, but transcription is unreliable for languages lacking strong speech recognisers, and it discards the tone and emotion that often carry the abuse itself. This paper examines whether abusive speech can instead be detected directly from audio, using CLAP, a model that learns a shared representation of sound and language, evaluated across ten Indic languages in the ADIMA dataset. A lightweight classifier trained on CLAP's existing audio representations, without adapting the model itself, comes within one to three points of a fully supervised system, and far outperforms prompting with no labelled examples at all. Further adaptation with a handful of labelled examples per language yields little extra benefit, varying unpredictably across languages. CLAP-based audio representations thus already offer a strong, inexpensive foundation for detecting abusive speech across languages, lowering the labelled data needed in practice.

Figures & tables

Figure 1: Overview of the proposed CLAP-based pipeline: a $k$ -shot subset adapts CLAP via supervised contrastive learning (projection-only, or projection+fine-tuning), and the resulting embeddings are evaluated under cross-lingual and leave-one-language-out settings.

Language	ADIMA (0-shot)	Projection-only	Projection+FT
Bengali	79.10	76.3 (0-shot)	76.3 (0-shot)
Bhojpuri	–	70.5 (5-shot)	70.5 (50-shot)
Gujarati	–	74.6 (0-shot)	74.6 (0-shot)
Haryanvi	–	77.6 (10-shot)	77.4 (0-shot)
Hindi	80.70	77.8 (0-shot)	77.8 (0-shot)
Kannada	78.40	76.8 (0-shot)	76.8 (0-shot)

Table 1: Cross-lingual macro-F1 (%) on ADIMA. Projection-only and projection+fine-tuning report mean macro-F1 over 5 support-set draws at the best-performing shot size (single run at

k=0

), shown in parentheses; ADIMA results from Gupta et al. (2022) shown where available. Best score per row in bold.

Panel of Figure 2 — Figure 2: LOLO vs. cross-lingual evaluation. Top: per-language macro-F1 difference (positive = LOLO stronger). Bottom: mean macro-F1 by shot size.

Encoder	0-shot	best few-shot	avg
CLAP ( clap-htsat-unfused )	76.6	75.9 (1-shot)	75.9
Whisper Large-v3	77.1	79.2 (50-shot)	78.4
Wav2Vec2-large	75.6	74.4 (1-shot)	74.3
XLS-R-300m †	76.2	75.3 (50-shot)	75.1
Wav2Vec2-base	77.2	76.0 (1-shot)	76.0

Table 2: Encoder ablation, projection-only (cross-lingual mean macro-F1 over 5 draws for

k\geq 1

; single run at

k=0

). “avg” is the mean over all six shot sizes.

\dagger

: XLS-R’s pre-training includes 7 of the 10 target languages, so cannot cleanly isolate the SSL objective from prior exposure.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Language	1-shot	5-shot	10-shot	25-shot	50-shot	0-shot
Bengali	74.8 $\pm$ 0.3	75.0 $\pm$ 2.3	74.5 $\pm$ 0.9	74.8 $\pm$ 1.3	74.2 $\pm$ 1.3	76.3
Bhojpuri	70.0 $\pm$ 0.9	70.5 $\pm$ 0.7	69.9 $\pm$ 1.4	70.2 $\pm$ 0.8	69.3 $\pm$ 1.0	68.9
Gujarati	74.1 $\pm$ 1.0	73.7 $\pm$ 0.9	74.6 $\pm$ 0.8	73.7 $\pm$ 1.6	73.6 $\pm$ 0.7	74.6
Haryanvi	77.0 $\pm$ 1.2	77.2 $\pm$ 1.8	77.6 $\pm$ 0.8	76.8 $\pm$ 0.6	77.4 $\pm$ 1.0	77.4
Hindi	77.2 $\pm$ 0.3	75.6 $\pm$ 0.9	76.4 $\pm$ 1.0	77.3 $\pm$ 1.1	76.8 $\pm$ 0.8	77.8
Kannada	75.6 $\pm$ 0.4	75.7 $\pm$ 0.6	76.0 $\pm$ 1.4	75.6 $\pm$ 0.3	74.7 $\pm$ 1.0	76.8

Appendix

Table 3: Cross-lingual macro-F1 (mean

\pm

std over 5 seeds), projection-only.

k{=}0

is a single deterministic run (no support-set sampling).

Language	Best shot (proj-only, mean)	0-shot optimal?
Bengali	0	yes
Bhojpuri	5	no
Gujarati	0	yes
Haryanvi	10	no †
Hindi	0	yes
Kannada	0	yes

Appendix

Table 4: Best-performing shot size per language (cross-lingual, projection-only; Table 3 ).

\dagger

: gain within one std.

\ddagger

: gain approximately one std.

Prompt variant	mean MF1	min	max
P1_basic	53.6	45.1	60.2
P2_hate	65.6	62.3	73.6
P3_offensive	65.2	59.8	72.6
P4_aggressive	64.3	58.4	75.1

Appendix

Table 5: Zero-shot prompt sensitivity (macro-F1 across the ten languages); min/max taken across languages per prompt variant.

Language	Macro-F1 (mean)	Best shot
Bengali	76.3	0-shot
Bhojpuri	70.6	5-shot
Gujarati	74.6	0-shot
Haryanvi	77.7	10-shot
Hindi	77.8	0-shot
Kannada	76.8	0-shot

Appendix

Table 6: Best LOLO results, projection-only (mean macro-F1 over 5 draws,

k\geq 1

; single run at

k=0

Language	Macro-F1 (mean)	Best shot
Bengali	76.3	0-shot
Bhojpuri	70.2	10-shot
Gujarati	74.6	0-shot
Haryanvi	77.4	0-shot
Hindi	77.8	0-shot
Kannada	76.8	0-shot

Appendix

Table 7: Best LOLO results, projection+fine-tuning (mean macro-F1 over 5 draws,

k\geq 1

; single run at

k=0

Figure 3: Language-wise LOLO vs. cross-lingual macro-F1, projection-only and projection+fine-tuning.

Hate Speech Detection

Hate Speech Detection identifies hateful or abusive language, its targets and intent, to support safer and more accountable content moderation.

See all 100 papers

Drift-Augmented Scoring: Text-Derived Noise Robustness for Zero-Shot Audio-Language Classification
Jun 3, 2026Tu Vo, Sheir Zaheer, Chan Y. ParkFew-Shot Audio Classification Audio Understanding
KC Machine Learning Lab
When Spoof Detectors Travel: Evaluation Across 66 Languages in the Low-Resource Language Spoofing Corpus
Mar 2, 2026Kirill Borodin, Vasiliy Kudryavtsev, Maxim Maslov +2Audio Deepfake Detection Speech Processing
lab260, Moscow Technical University of Communications and Informatics, Russia · BitmanagerAI, UAE
Transductive Zero-Shot Audio Classification with Audio-Language Models
Jun 15, 2026Jingwen Zhou, Mingzhe WangAudio-Language Model Evaluation Few-Shot Audio Classification
Xidian University, Xi’an, China

Jun 3, 2026J/K move · Enter open · S save

Drift-Augmented Scoring: Text-Derived Noise Robustness for Zero-Shot Audio-Language Classification

Tu Vo, Sheir Zaheer, Chan Y. Park

KC Machine Learning Lab

Few-Shot Audio Classification Audio Understanding Audio-Language Models Noise Robustness Audio Classification

Open paper AskPDF arXiv

Contrastive audio-language models such as CLAP enable zero-shot audio classification: a sound is labelled by matching its embedding to text prompt embeddings, with no labelled audio. This matching breaks down under acoustic noise, where accuracy and mAP fall by 12-30 percentage points at 0 dB SNR on standard benchmarks. We propose Drift Augmented Scoring (DAS), a small per-class bonus added to the cosine score. The bonus rewards a class when the noisy audio embedding drifts in the direction that the class's noise-conditioned text prompts predict. It is derived from text alone, computed once and cached, and adds a single inner product per class at inference, with no gradients and no test-time batch. On a LAION CLAP backbone, we compare DAS against the four variants of Acevedo et al.'s concurrent method on UrbanSound8K and the full FSD50K eval set, mixing each clip with urban acoustic scene noise across a range of SNRs. DAS improves the metric on every test condition: by +2.60 to +5.75 accuracy points on UrbanSound8K and +1.50 to +1.74 mAP points on FSD50K.

Abstract

Hate Speech Detection

Explore similar work

Drift-Augmented Scoring: Text-Derived Noise Robustness for Zero-Shot Audio-Language Classification

When Spoof Detectors Travel: Evaluation Across 66 Languages in the Low-Resource Language Spoofing Corpus

Transductive Zero-Shot Audio Classification with Audio-Language Models