High Performance Personal Adaptation Speech Recognition Framework by Incremental Learning with Plural Language Models

This paper introduces a speech recognition framework for high performance personalized adaption. It is based on plural language models and personalized incremental learning interface for error correction. If an error in a recognition result is detected by a bidirectional neural language model, it generates a corrected sentence by a majority decision among multiple n-gram language models considering several aspects. Moreover, we introduce a speaker adaptation by updating language models through incremental learning, which can adjust the parameter from training data. The experiments show that our framework improves word-error rate to 78% compared with Google Chrome's Speech Recognition API. Our framework can be used for improving one-to-one human-machine dialogue systems such as intelligent (counseling) agents.