Development of highly accurate real-time large scale speech recognition system

This paper describes the development of the framework and the algorithm for large scale automatic speech recognition systems. Technical advances include the acceleration of decoding speed by leveraging the computational power of many-core graphic processing units (GPU), in order to solve the issue of training data sparseness, improvement in the accuracy by Subspace Gaussian Mixture Models (SGMM), and employing novel methods of language models such as the Instant Language Model Adaptation (ILMA) method. We present the effectiveness of each technique by evaluating it with actual usage data collected from television sets. It is shown that the proposed engine can recognize speech at real time with high accuracy.