| Summary: | 碩士 === 國立臺北護理健康大學 === 語言治療與聽力研究所 === 105 === The present study aimed to evaluate the effects of a new noise reduction (NR) algorithm based on convolutional neural network (CNN), as compared to other algorithms, on speech recognition performance and listening efforts.
A speech recognition test and a listening effort scale were conducted in various listening conditions to evaluate whether improved speech recognition in noise and less listening efforts expended were evident. Thirty adults with normal hearing participated in the present study. For the speech recognition test, lists of sentences from Taiwan Mandarin Hearing in Noise Test (TMHINT) were first mixed with noise and were then processed using the NR algorithms. Participants were asked to listen to the unprocessed and processed sentences and were required to repeat heard sentences as verbatim as possible. Participants also rated the extents to which the efforts expended for them to recognize the sentences in different listening conditions.
Sentences processed using the CNN-match algorithm has resulted in significantly better speech recognition scores among all NR algorithms. There is, however, no significant difference between the CNN-match condition and the unprocessed condition. On the other hand, listening effort ratings showed that the CNN-match condition was the most preferable among all listening conditions.
Sentences processed using the CNN-based NR algorithm has resulted in speech recognition scores equal to but not worse than the unprocessed sentences. Least efforts reported by the participants listening to the sentences processed with the CNN-based NR algorithm among all other listening conditions demonstrated the inevitable role of listening efforts in evaluating speech recognition ability.
|
|---|
