| Summary: | 碩士 === 元智大學 === 電機工程學系 === 106 === Hearing is one of the most direct and efficient methods of human speech communication. While people are suffering from hearing loss, it will directly influence their living quality and bring about a series of related social problems. According to estimates by the World Health Organization, the current hearing impaired population has reached 466 million, and WHO estimates that the hearing impaired population will reach 900 million by 2050. This trend directly indicates that hearing loss-related research requires more attention to improve the hearing quality of hearing impaired people.
For hearing loss, the most common solution is to wear hearing aids. Although hearing aids have enhanced the quality of life of hearing impaired people in quiet conditions, there is still room for improvement in noisy environments. To effectively solve this problem, the noise reduction algorithm was used in the development of hearing aids. More recently, a noise reduction method based on deep learning architecture called deep denoising autoencoder (DDAE) was proposed. According to previous research, DDAE provides better listening quality than current noise reduction methods for normal hearing people, but DDAE has not been applied effectively to hearing impaired people. Therefore, this paper explores the processing benefits of the DDAE noise reduction method for a variety of typical hearing loss types. After an objective evaluation, we prove that deep learning architecture provides better hearing benefits for the hearing impaired. In addition, we combined the concept of multi-task learning with DDAE architecture to further improve the performance of DDAE NR model, called M-DDAE. This new architecture improved the listening performance of the DDAE architecture for hearing loss in noisy situations.
In addition to the noise reduction method, an appropriate auditory compensation unit is necessary for the patient to provide sufficient listening ability. Currently, most of the hearing aids' compensation structures use nonlinear gain adjustment based on the input speech volume for the patient. Therefore, this paper also proposes a speech energy classifier based on the deep neural network architecture.
In the future, we will combine two proposed architectures to try to improve the compensation benefits for hearing impaired patients.
|
|---|
