PulseAugur / Brief
EN
LIVE 06:35:38

Brief

last 24h
[2/2] 221 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

  1. Zero-Shot Parkinson's Disease Detection from Speech: Comparing Large Audio and Language Models

    Researchers have compared two methods for using large language and audio models to detect Parkinson's disease from speech without prior training. The study found that performance varied depending on whether the models processed handcrafted acoustic features or raw audio waveforms. While handcrafted features offered more consistent results in low-resource languages like Bengali, direct audio input showed dataset-dependent improvements. AI

    IMPACT Investigates how different AI model input modalities affect performance in zero-shot disease detection from speech.

  2. Polite on the Surface, Wrong in Practice: A Curated Dataset for Fixing Honorific Failures in Multilingual Bangla Generation

    Researchers have developed a new dataset and benchmarking framework called BLADE to address honorific failures in multilingual Bangla text generation. This dataset comprises over 4,000 curated interaction pairs designed to improve the cultural nuance and context-dependent communication of large language models. Fine-tuning models like DeepSeek-8B and LLaMA-3.2-3B on BLADE has shown significant improvements in structural fidelity and honorific alignment for low-resource languages. AI

    IMPACT Enhances multilingual LLM capabilities by addressing cultural nuances and honorifics in low-resource languages like Bangla.