| Summary: | 博士 === 國立臺灣大學 === 電信工程學研究所 === 100 === Multimedia content over the Internet is very attractive, while the spoken part of such content very often tells the core information. Therefore, spoken content retrieval will be very important in helping users retrieve and browse efficiently across the huge qualities of multimedia content in the future. There are usually two stages in typical spoken content retrieval approaches. In the first stage, the audio content is recognized into text symbols by an Automatic Speech Recognition (ASR) system based on a set of acoustic models and language models. In the second stage, after the user enters a query, the retrieval engine searches through the recognition output and returns to the user a list of relevant spoken documents or segments. If the spoken content can be transcribed into text with very high accuracy, the problem is naturally reduced to text information retrieval. However, the inevitable high recognition error rates for spontaneous speech under a wide variety of acoustic conditions and linguistic context make this never possible. In this thesis, the above standard two-stage architecture is completely broken, and the two stages of recognition and retrieval are mixed up and considered as a whole. A set of approaches beyond retrieving over recognition output has been developed here. This idea is very helpful for spoken content retrieval, and may become one of the main future directions in this area.
To consider the two stages of recognition and retrieval as a whole, it is proposed to adjust the acoustic model parameters borrowing the techniques of discriminative training but based on user relevance feedback. The problem of retrieval oriented acoustic model re-estimation is different from the conventional acoustic model training approaches for speech recognition in at least two ways:
1. The model training information includes only whether a spoken segment is relevant to a query or not; it does not include the transcription of any utterance. 2. The goal is to improve retrieval performance rather than recognition accuracy. A set of objective functions for retrieval oriented acoustic model re-estimation is proposed to take the properties of retrieval into consideration.
There have been some previous works in spoken content retrieval taking advantage of the discriminative capability of machine learning methods. Different from the previous works that derive information from recognition output as features, acoustic vectors such as MFCC are taken as the features for discriminating relevant and irrelevant segments, and they are successfully applied on the scenario of Pseudo Relevance Feedback (PRF).
The recognition process can be considered as ``quantization'', in which the acoustic vector sequences are quantized into word symbols. Because different vector sequences may be quantized into the same symbol, much of the information in the spoken content may be lost in the stage of speech recognition. Information directly from the acoustic vector space is considered to compensate for the recognition output in this thesis. This is realized by either PRF or a graph-based re-ranking approach considering the similarity structure among all the segments retrieved. This approach is successfully applied on not only word-based retrieval system but also subword-based system, and these approaches improve the results of Out-of-Vocabulary (OOV) queries as well.
The task of Spoken Term Detection (STD) is mainly considered in this thesis, for which the goal is simply returning spoken segments that contain the query terms. Although most works in spoken content retrieval nowadays continue to focus on STD, in this thesis a more general task is also considered: to retrieve the spoken documents semantically related to the queries, no matter the query terms are included in the spoken documents or not. Taking ASR transcriptions as text, the techniques such as latent semantic analysis or query expansion developed for text-based information retrieval can be directly applied for this task. However, the inevitable recognition errors in ASR transcriptions degrade the performance of these techniques. To have more robust semantic retrieval of spoken documents, the expected term frequencies derived from the lattices are enhanced by acoustic similarity with a graph-based approach. The enhanced term frequencies improve the performance of language modelling retrieval approach, document expansion techniques based on latent semantic analysis, and query expansion methods considering both words and latent topic information.
|
|---|
