Reducing the OOV rate in broadcast news speech recognition

Thomas Kemp Alex Waibel Interactive Systems Laboratories, ILKD University of Karlsruhe 76128 Karlsruhe, Germany ABSTRACT The recognition of broadcast news is a challenging problem in speech recognition. To achieve the long-term goal of robust, real-time news transcription, several problems have to be overcome, e.g. the variety of acoustic conditions and the unlimited vocabulary. In this paper we address the problem of unlimited vocabulary. We show, that this problem is more serious for German than it is for English. Using a speech recognition system with a large vocabulary, we dynamically adapt the active vocabulary to the topic of the current news segment. This is done by using information retrieval (IR) techniques on a large collection of texts automatically gathered from the internet. The same technique is also used to adapt the language model of the recognition system. The process of vocabulary adaptation and language model retraining is completely unsupervised. We show, that dynamic vocabulary adaptation can significantly reduce the out-of-vocabulary (OOV) rate and improve the word error rate of our broadcast news transcription system View4You. 1. THE VIEW4YOU SYSTEM The View4You project is a cooperation between the Interactive Systems Labs and the Carnegie Mellon University's Informedia group [5]. It aims at the automatic generation of a searchable multilingual video database. In the prototype system, German and Serbocroatian TV news shows are recorded daily and stored as MPEG compressed les. Using the acoustic signal, a segmenter chops the newscasts into acoustically homogeneous segments ranging from several seconds to few minutes in length. A speech recognition system generates transcriptions for the segments. The segmentation information and the automatic transcriptions are stored in a database. The user of the system can give queries in natural language, e.g. 'Tell me everything about the peace talks between Mr Netanyahu and Mr Arafat'. Using the speech recognizer's transcriptions in the multimedia database, an information retrieval component computes a ranked order of relevant segments, which are displayed to the user. By clicking on a segment, an MPEG-player is activated that plays the corresponding video segment. For more details on the View4You system, see [1]. 2. MOTIVATION The index into View4You's video database consists of the output of our speech recognizer. Therefore, only words that are in the vocabulary of the recognizer can be searched for. If a video contains keywords that are unknown to the recognizer, they cannot be found in the index, and the user can not retrieve the video by this keyword. OOV (out-of-vocabulary) words therefore pose a problem to the View4You system, and the vocabulary of the speech recognizer should be as large as possible to ensure low OOV rates. Currently, our speech recognizer is limited to a vocabulary of 64k words. On the North American Business News (NAB) corpus, a vocabulary of this size covers more than 99% of the text, and even with a 20k vocabulary, OOV rates on the NAB task do not exceed 3%. We measured the OOV rate on German news shows for a 60k vocabulary which was derived from our language model corpus. The result is shown in gure 1.