Research Direction of Speech Recognition Technology

The difficulty in doing good speech recognition in noisy environments is how to separate the noise from the human voice. Traditional audio recognition requires manually designed modules and relies on Hidden Markov Models, which often requires a lot of manpower and experience to adjust the model noise and speech variants. The main research direction for the future is to replace Hidden Markov Models with deep learning, such as deep neural networks (DNNs) based on recurrent neural networks for acoustic modeling, making speech recognition systems much simpler. Hitachi claims to have developed a new technology that separates noise from speech by taking advantage of the fact that the volume of the conversation varies less than the noise.