Multistage Spoken Term Detection for Searching Speech Databases

Abstract

The availability of high-speed, low-cost Internet enables the use of multimedia files for a variety of applications in our day-to-day activities. A considerable volume of audio archives is available on different websites. Such vast audio resources are useful only newlineif the needed file can be retrieved accurately and efficiently. Audio search refers to the search and retrieval of a particular audio file from a large audio database. Since most of newlinethe audio archives are not well indexed or labelled, it is still a challenging task. newlineSpoken Term Detection (STD) refers to the process of locating the occurrences of newlinespoken queries in a large speech database. Generally, two methods have been adopted newlinefor STD: an Automatic Speech Recognition (ASR) based label sequence matching newlineor feature-based template matching. ASR-based techniques utilize phoneme models newlineof a language, which require a considerable amount of labelled training data in the newlineselected language. Hence such techniques are considered as language-dependent, and newlineit is not feasible to develop ASR for each language. The feature-based template newlinematching techniques address this task in a language-independent manner, but they are newlinecomputationally complex. This work combines the positive aspects of both the methods newlineby introducing a multistage architecture to address the task of STD for low-resourced newlinelanguages. Two different approaches have been proposed for Language-Dependent (LD) newlineand Language-Independent (LI) STD. newlineFor Language-Dependent STD (LD-STD), a Phoneme Recognizer (PR) trained for newlinea particular language is used to convert the input speech into a corresponding phoneme newlinesequence. After converting the input speech to corresponding text sequences, two stages newlinenamed coarse and fine search stages help to spot the query locations. In coarse search, newlinelabel level matching is performed using a sequence matching technique for shortlisting newlinethe probable query locations. A customized local alignment technique is proposed for newlinesequence matching.