Lattice segmentation techniques developed for Minimum Bayes Risk decoding in large vocabulary speech recognition tasks are used to compute the statistics for discriminative training algorithms that estimate HMM parameters so as to reduce the overall risk over the training data. New estimation procedures are developed and evaluated for small vocabulary and large vocabulary recognition tasks, and additive performance improvements are shown relative to maximum mutual information estimation. These relative gains are explained through a detailed analysis of individual word recognition errors.

This article describes approaches to content-based access to spoken audio with a qualitative and tutorial emphasis. We describe how the analysis, retrieval and delivery phases contribute making spoken audio content more accessible, and we outline a number of outstanding research issues. We also discuss the main application domains and try to identify important issues for future developments. The structure of the article is based on general system architecture for content-based 2 access which is depicted in Figure 1. Although the tasks within each processing stage may appear unconnected, the interdependencies and the sequence with which they take place vary

In this paper, a cross-media browsing demonstrator named InfoLink is described. InfoLink automatically links the content of Dutch broadcast news videos to related information sources in parallel collections containing text and/or video. Automatic segmentation, speech recognition and available meta-data are used to index and link items. The concept is visualised using SMIL-scripts for presenting the streaming broadcast news video and the information links. 1.

The application of automatic speech recognition in the broadcast news domain is well studied. Recognition performance is generally high and accordingly, spoken document retrieval can successfully be applied in this domain, as demonstrated by a number of commercial systems. In other domains, a similar recognition performance is hard to obtain, or even far out of reach, for example due to lack of suitable training material. This is a serious impediment for the successful application of spoken document retrieval techniques for other data then news. This paper outlines our first steps towards a retrieval system that can automatically be adapted to new domains. We discuss our experience with a recently implemented spoken document retrieval application attached to a web-portal that aims at the disclosure of a multimedia data collection in the oral history domain. The paper illustrates that simply deploying an off-theshelf broadcast news system in this task domain will produce error rates that are too high to be useful for retrieval tasks. By applying adaptation techniques on the acoustic level and language model level, system performance can be improved considerably, but additional research on unsupervised adaptation and search interfaces is required to create an adequate search environment based on speech transcripts. 1.

Abstract. This paper discusses audio indexing tools that have been implemented for the disclosure of Dutch audiovisual cultural heritage collections. It explains the role of language models and their adaptation to historical settings and the adaptation of acoustic models for homogeneous audio collections. In addition to the benefits of cross-media linking, the requirements for successful tuning and improvement of available tools for indexing the heterogeneous A/V collections from the cultural heritage domain are reviewed. And finally the paper argues that research is needed to cope with the varying information needs for different types of users. 1

The task for the CLEF-2005 cross-language speech retrieval track was to identify topically coherent segments of English interviews in a known-boundary condition. Seven teams participated, performing both monolingual and cross-language searches of ASR transcripts, automatically generated metadata, and manually generated metadata. Results indicate that monolingual search technology is sufficiently accurate to be useful for some purposes (the best mean average precision was 0.18) and cross-language searching yielded results typical of those seen in other applications (with the best systems approximating monolingual mean average precision). 1.

Access to historical audio collections is typically very restricted: content is often only available on physical (analog) media and the metadata is usually limited to keywords, giving access at the level of relatively large fragments, e.g., an entire tape. Many spoken word heritage collections are now being digitized, which allows the introduction of more advanced search technology. This paper presents an approach that supports online access and search for recordings of historical speeches. A demonstrator has been built, based on the so-called Radio Oranje collection, which contains radio speeches by the Dutch Queen Wilhelmina that were broadcast during World War II. The audio has been aligned with its original 1940s manual transcriptions to create a time-stamped index that enables the speeches to be searched at the word level. Results are presented together with related photos from an external database. 1

Rapid and inexpensive techniques for automatic transcription of speech have the potential to dramatically expand the types of content to which information retrieval techniques can be productively applied, but limitations in accuracy and robustness must be overcome before that promise can be fully realized. Combining retrieval results from systems built on various errorful representations of the same collection offers some potential to address these challenges. This paper explores that potential by applying Generalized Additive Models to optimize the combination of ranked retrieval results obtained using transcripts produced automatically for the same spoken content by substantially different recognition systems. Topic-averaged retrieval effectiveness better than any previously reported for the same collection was obtained, and even larger gains are apparent when using an alternative measure emphasizing results on the most difficult topics. 1

This paper introduces a new approach to ranking speech utterances by a system’s confidence that they contain a spoken word. Multiple alternate pronunciations, or degradations, of a query word’s phoneme sequence are hypothesized and incorporated into the ranking function. We consider two methods for hypothesizing these degradations, the best of which is constructed using factored phrasebased statistical machine translation. We show that this approach is able to significantly improve upon a state-of-the-art baseline technique in an evaluation on held-out speech. We evaluate our systems using three different methods for indexing the speech utterances (using phoneme, phoneme multigram, and word recognition), and find that degradation modeling shows particular promise for locating out-of-vocabulary words when the underlying indexing system is constructed with standard word-based speech recognition. 1

This paper briefly reviews state of the art related to the topic of speech variability sources in automatic speech recognition systems. It focuses on some variations within the speech signal that make the ASR task difficult. The variations detailed in the paper are intrinsic to the speech and affect the different levels of the ASR processing chain. For different sources of speech variation, the paper summarizes the current knowledge and highlights specific feature extraction or modeling weaknesses and current trends. 1.