Low Resource Efficient Speech Retrieval

Speech retrieval;automatic speech recognition;keyword search;spoken document classification;Electrical Engineering

Liu, ChunxiKhudanpur, Sanjeev ;

Johns Hopkins University

关键词: Speech retrieval;    automatic speech recognition;    keyword search;    spoken document classification;    Electrical Engineering;   

Others  :  https://jscholarship.library.jhu.edu/bitstream/handle/1774.2/60126/LIU-DISSERTATION-2018.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: JOHNS HOPKINS DSpace Repository

【 摘 要 】

Speech retrieval refers to the task of retrieving the information, which is useful or relevant to a user query, from speech collection. This thesis aims to examine ways in which speech retrieval can be improved in terms of requiring low resources - without extensively annotated corpora on which automated processing systems are typically built - and achieving high computational efficiency. This work is focused on two speech retrieval technologies, spoken keyword retrieval and spoken document classification. Firstly, keyword retrieval - also referred to as keyword search (KWS) or spoken term detection - is defined as the task of retrieving the occurrences of a keyword specified by the user in text form, from speech collections. We make advances in an open vocabulary KWS platform using context-dependent Point Process Model (PPM). We further accomplish a PPM-based lattice generation framework, which improves KWS performance and enables automatic speech recognition (ASR) decoding.Secondly, the massive volumes of speech data motivate the effort to organize and search speech collections through spoken document classification. In classifying real-world unstructured speech into predefined classes, the wildly collected speech recordings can be extremely long, of varying length, and contain multiple class label shifts at variable locations in the audio. For this reason each spoken document is often first split into sequential segments, and then each segment is independently classified. We present a general purpose method for classifying spoken segments, using a cascade of language independent acoustic modeling, foreign-language to English translation lexicons, and English-language classification. Next, instead of classifying each segment independently, we demonstrate that exploring the contextual dependencies across sequential segments can provide large classification performance improvements. Lastly, we remove the need of any orthographic lexicon and instead exploit alternative unsupervised approaches to decoding speech in terms of automatically discovered word-like or phoneme-like units. We show that the spoken segment representations based on such lexical or phonetic discovery can achieve competitive classification performance as compared to those based on a domain-mismatched ASR or a universal phone set ASR.