bangla

The paper introduces BanglaTurn, a benchmark corpus and Whisper-based model for end-of-turn detection in Bangla speech, achieving 84.33% accuracy compared to a 69.28% baseline.

BanglaKontho presents a 20-hour single-speaker Bangla TTS corpus from professional audiobooks, along with an open-source text normalizer and preprocessing pipeline, to address the gap in long-form prosody for low-resource Bangla speech synthesis.

This paper reports on training a language model end-to-end in Rust, detailing failures in Rust ML frameworks like Candle and Burn, and proposing verification methods, with the conclusion that Rust is currently better suited for model serving than training.

This paper proposes HybridRAG-BN, a retrieval-augmented framework for Bangla knowledge-base question answering that combines hybrid retrieval, Gemma-based generation, and LoRA fine-tuned verification, achieving first place with F1 scores of 0.71654 and 0.72912.

This paper presents a unified poly-dialectal neural machine translation system for 12 Bangla regional dialects, introducing the largest multi-dialect parallel corpus to date and achieving state-of-the-art BLEU scores with a fine-tuned BanglaT5 model using DoRA.

This arXiv paper audits five frontier LLMs on native Bangla derogatory speech, finding that safety alignment fails to generalize to low-resource languages — models comprehend and generate unsafe content at high rates despite high-resource alignment. The authors propose a 'comprehension–containment decoupling' and show that reasoning and persona framing further break down safety filters.

Introduces Khondo, the first benchmark for document packet splitting on Bangladeshi government forms. It is a vision-native, bilingual dataset spanning multiple concatenation schemes, and zero-shot evaluation of MLLMs reveals that page-order reconstruction remains a key open challenge.

This paper introduces a benchmark dataset of 1,516 expert-verified Bangla sentences for disambiguating culturally entangled homographs (words that are both names and common nouns). It shows that LLMs suffer from dominant-meaning bias and proposes contrastive chain-of-thought prompting and distillation to reduce this bias.

Introduces BaFCo, a benchmark dataset for Bangla form comprehension focusing on Document Layout Analysis (DLA) and Key Information Extraction (KIE). It includes 200 multi-page complex Bangladeshi government forms with fine-grained annotations across 26 entity types and evaluates multiple MLLMs, revealing limitations in understanding complex Bangla forms.

This paper introduces a Bangla event detection benchmark with noisy text (ASR, orthographic corruption) and evaluates encoder-only and decoder-only LLMs, finding decoder models more robust to noise.

This paper introduces BLADE, a culturally aligned instruction-tuning dataset of 4,196 interaction pairs for fixing honorific failures and pragmatic gaps in multilingual Bangla generation. Fine-tuning models like DeepSeek-8B and LLaMA-3.2-3B on this dataset yields substantial improvements in structural fidelity and honorific alignment.