Ke Hu

Title: The Innovative Contributions of Ke Hu

Introduction

Ke Hu is a prominent inventor based in Stony Brook, NY, known for his significant contributions to the field of technology. With a total of seven patents to his name, he has made remarkable advancements in the area of speech recognition and deliberation models. His work is characterized by a blend of innovative methodologies and practical applications.

Latest Patents

Among his latest patents, Ke Hu has developed a method for deliberation by text-only and semi-supervised training. This method involves receiving training data that includes unspoken textual utterances, which are not paired with any corresponding spoken utterance of non-synthetic speech. The process trains a deliberation model that consists of a text encoder and a deliberation decoder on these unspoken textual utterances. Additionally, the model receives first-pass hypotheses and non-causal acoustic embeddings, generated by a recurrent neural network-transducer (RNN-T) decoder. The method further encodes the first-pass hypotheses and generates second-pass hypotheses using the deliberation decoder.

Another notable patent is the transducer-based streaming deliberation for cascaded encoders. This method includes receiving a sequence of acoustic frames and generating a first higher-order feature representation for each corresponding acoustic frame. It also involves generating a first pass speech recognition hypothesis and a text encoding for that hypothesis. The method continues with generating a second higher-order feature representation and a second pass speech recognition hypothesis, enhancing the accuracy and