Cross-lingual Embedding Clustering for Hierarchical Softmax in Low-Resource Multilingual Speech Recognition
Researchers have developed a new method for low-resource multilingual speech recognition by employing cross-lingual embedding clustering to create a hierarchical Softmax decoder. This technique improves upon previous methods by assessing token similarity more deeply, leading to enhanced accuracy in multilingual Automatic Speech Recognition (ASR) systems, particularly for languages with limited data. Experiments on a 15-language dataset validated the approach's effectiveness. AI