SN-WER: Script-Normalized WER for Multi-Script Indic ASR Evaluation
Two new research papers introduce methods to improve automatic speech recognition (ASR) systems. The first paper, "MURMUR," presents an efficient inference system designed for long-form ASR that balances accuracy and low latency by processing audio in intermediate-sized chunks and optimizing attention sparsity. The second paper, "SN-WER," proposes a new evaluation metric called Script-Normalized WER (SN-WER) to address inaccuracies in standard Word Error Rate calculations when dealing with multilingual ASR, particularly for Indic languages, by normalizing scripts before comparison. AI
IMPACT These papers introduce novel techniques for improving ASR accuracy and evaluation, potentially leading to more robust speech-to-text systems for diverse languages and long-form content.