پياده‌سازي سيستم‌هاي تركيبي حذف انعكاس از گفتار و ارايه دادگان دوكاناله گفتار فارسي براي سنجش كارآيي آنها

Implementation of Hybrid Speech Dereverberation Systems and Proposing Dual Microphone Farsi Database in Order to Evaluating Enhancement Systems

در اين مقاله جوانب مختلف مساله حذف انعكاس از سيگنال گفتارفارسي بررسي مي‌شود. به عنوان مصالحه‌اي بين ميزان بهسازي گفتار از يكسو و پيچيدگي و بار محاسباتي روش از سوي ديگر، تعداد دو ميكروفون در سيستم در نظر گرفته شده است. به منظور امكان سنجش و مقايسه كارآيي سيستم‌هاي مورد نظر، ابتدا مجموعه لغات فارسي براي آزمون قابليت فهم گفتار تهيه و سپس با استفاده از آن، دادگان دو كاناله گفتار فارسيِ داراي انعكاس طراحي و ضبط گرديد. در اين تحقيق، روش‌هاي پايه (تك‌كاناله و دوكاناله) حذف انعكاس از گفتار مطالعه و پياده سازي و با استفاده از دادگان تهيه شده مقايسه گرديده‌اند. بر پايه اين مطالعه، يك سري از روش‌هاي تركيبي كه قابليت فهم و كيفيت سيگنال گفتار آغشته به انعكاس را به نحو مطلوبي افزايش مي‌دهد، ارايه شده‌است و نشان داده شده است كه تركيب سه روش delay-and-sum beamforing، فيلتر معكوس، و روش مبتني بر خطاي فاز نسبت به ساير سيستم‌هاي تركيبي از لحاظ ميزان كيفيت و قابليت فهم گفتار خروجي منجر به نتايج بهتري مي‌شود.

In various applications, such as speech recognition and automatic teleconferencing, the recorded speech signals may be corrupted by both noise and reverberation. Reverberation causes a noticeable change in speech intelligibility and quality. In this research, firstly reverberation is described. There are some de-reverberation enhancement algorithms that use only one microphone. They mostly use inverse filtering and spectral subtraction as their sub-systems. On the other hand, there are many multi-microphone speech enhancement systems; Delay-and-sum beam former is the most famous amongst them. Moreover, several efficient approaches have been also reported that use linear prediction (LP) residual signal, inverse filtering, and phase error. Despite the improvements and benefits gained by the use of several input microphones, considering the tradeoff between these gains and the complexity and computational cost forced by the use of more microphones, many researchers have focused on dual-microphones systems. So, a review on Microphone array signal processing is explained and then an arrangement for two microphones systems is proposed. As we want to evaluate these algorithms for Farsi speech signals, the problem of speech intelligibility assessment has been explained and a Farsi word list for Diagnostic Rhyme Test (DRT) is presented.The structure of presented word list is similar to that of English DRT words. In this research, after a brief study of above-mentioned methods, we propose and implement some hybrid techniques to benefit from the advantages of several methods and achieve significant improvement in output signals. It will be shown that the proposed method performs superior to the state-of-the-art dereverberation algorithms.