CUET_InferX@DravidianLangTech 2026: Shared Task on Dialect Based Speech Recognition and Classification in Tamil

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

The CUET_InferX team developed a system for the DravidianLangTech@ACL 2026 Shared Task on Dialect Based Speech Recognition and Classification in Tamil, addressing challenges posed by Tamil's internal variability, including casual conversations, code mixing, and regional phonetic differences. For Subtask 2, Automatic Speech Recognition (ASR), their dual-architecture system combined a fine-tuned Whisper-large-v3 model with Low-Rank Adaptation (LoRA) and a Wav2Vec2 XLSR-53 model, enhanced by a KenLM statistical language model for n-gram phonetic correction. This ASR system achieved a Word Error Rate (WER) of 0.54, securing 2nd position. For Subtask 1, Speech-Based Dialect Classification, the team utilized a text-based weighted ensemble of IndicBERT, MuRIL, XLM-RoBERTa, and TamilBERT models, which relied entirely on the ASR system's transcription outputs. This classification system attained a Macro F1 score of 0.22, placing 9th in its subtask.

Key takeaway

For NLP Engineers developing speech technologies for highly variable languages like Tamil, consider a multi-model approach. Your ASR system's accuracy is critical, as demonstrated by its direct impact on downstream dialect classification performance. You should explore dual-architecture ASR combining fine-tuned large models with statistical language models, and then utilize its output for text-based ensemble classifiers to improve overall system robustness.

Key insights

A dual-architecture ASR system and a text-based ensemble for dialect classification address Tamil's linguistic variability.

Principles

• Address language variability with specialized models.
• ASR performance directly impacts downstream text-based tasks.
• Ensemble methods can improve classification.

Method

For ASR, fine-tune Whisper-large-v3 with LoRA and Wav2Vec2 XLSR-53, then apply KenLM for phonetic correction. For dialect classification, use ASR output with a weighted ensemble of BERT-based models.

In practice

• Fine-tune Whisper-large-v3 with LoRA.
• Integrate statistical language models for ASR.
• Use ASR output for text-dependent tasks.

Topics

• Tamil Language Processing
• Automatic Speech Recognition
• Dialect Classification
• Whisper-large-v3
• Wav2Vec2 XLSR-53
• LoRA Fine-tuning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Voice of India: A Large-Scale Benchmark for Real-World Speech Recognition in India — Real-world Indic ASR requires benchmarks with unscripted speech, diverse languages, and flexible spelling.
• dialect2vec: A vector-based method for dialectal transcription of Portuguese from ALiB questionnaires — dialect2vec uses byte-level encoding of IPA sequences to model dialectal variation, outperforming subword models.
• Improving low-resource ASR using bilingual fine-tuning with language identification: a cross-linguistic evaluation — Bilingual fine-tuning enhances low-resource ASR, particularly when language identification is accurate or explicitly guided during inference.
• Pretrained self-supervised speech models can recognize unseen consonants — Pretrained self-supervised speech models effectively generalize to and accurately recognize typologically rare phonemes, such as click consonants.
• MURMUR: An Efficient Inference System for Long-Form ASR — Murmur efficiently balances long-form ASR accuracy and latency using a two-level inference system.
• Paza: Introducing automatic speech recognition benchmarks and models for low resource languages — Human-centered design and evaluation are crucial for effective speech models in low-resource, real-world environments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.