室內遠距離語音辨識實驗

Experiments on In-House Far-Field Speech Recognition

近年來，語音辨識的應用推出各種遠距離操作的系統，例如車用語音助理、智慧音箱等，在這些系統中遠距離語音辨識扮演著關鍵的角色。本文主要提出我們在智慧音箱裝置上的遠距離語音辨識相關實驗及成果，我們利用資料擴充方式、模擬遠距離語音及基於類神經網路聲學模型來降低字元錯誤率。在實驗部分，本文利用智慧音箱錄製了三種距離的平行測試語料，其中50cm情境語料可從13.31％降至8.41％，相對改善36.8％，而80cm情境語料從19.20％降至10.89％，相對改善43.2％。

In recent years, speech recognition applications have introduced a variety of remote operating systems, such as car voice assistants, smart speakers, etc. In these systems, far-field speech recognition plays a key role. This paper mainly presents our experiments and results on far-field speech recognition on smart speaker devices. We use data augmentation methods, simulated far-field speech and neural network-based acoustic models to reduce the character error rate (CER). In the experimental part, this paper recorded the parallel test corpus of three distances using the smart speaker. The 50cm situation corpus can be reduced from 13.31% to 8.41%, the relative improvement is 36.8%, and the 80cm situation corpus is reduced from 19.20% to 10.89% with relative improvement of 43.2%.