Brief

Researchers have introduced the Urdu Katib Handwritten Dataset (UKHD), the first offline dataset of historical Urdu handwritten text lines. This dataset aims to address the scarcity of resources for Urdu Handwritten Text Recognition (UHTR). The study also evaluated various CRNN-based models, identifying CNN-BGRU-CTC as the most effective for Urdu Katib Handwriting Recognition, achieving low character and word error rates. AI

Researchers have introduced a new paradigm for evaluating automatic speech recognition (ASR) systems that aims to improve upon existing metrics like Word Error Rate (WER) and Character Error Rate (CER). The proposed method incorporates a chosen metric to generate a Minimum Edit Distance (minED), which better correlates with human perception and accounts for linguistic and semantic information. This approach allows for a more nuanced study of transcription error severity from a human perspective. AI

Researchers have developed new methods for aligning large language models (LLMs) with user preferences. One approach, TKTO, focuses on text-to-speech systems, enabling data-efficient, token-level optimization to improve pronunciation accuracy and reduce errors. Another framework, POPI, addresses LLM personalization by separating the process into a preference summary generator and a response generator, allowing for user-specific outputs and reducing context overhead. AI