Speech Models Struggle with Diverse Pronunciations

A recent study reveals that state-of-the-art speech recognition models struggle with accurately transcribing street names, particularly for speakers with diverse linguistic backgrounds, showing a 39% average error rate. The research introduces two new benchmarks, SF Streets and US Streets, to evaluate named entity recognition in real-world scenarios. By employing a synthetic data generation technique called cross-lingual style transfer, the study demonstrates a potential 60% improvement in accuracy with fewer than 1,000 training samples. This highlights a critical gap in current speech models and suggests a viable path forward for enhancing their reliability in high-stakes applications.