Development of a Speech Recognition System for Spanish Broadcast News

One of the ASR applications is the generation of transcripts to facilitate searching through multi-media collections containing spoken data. Especially in the broadcast news domain ASR systems have been successfully deployed to index large collections of news. First of all because retrieval performed on ASR generated transcripts with an word-error rate (WER) under 50% gives resonable results [1] and second because ASR systems nowdays achieve high performances on broadcastnews data WER rate below 10% are no longer unusual [2][3]. In the MESH project[4]whose goal is to extract, compare and combine multimedia content (audio, video and text) from multiple news sources ASR modules for three different languages (Spanish, German and English) are going to be integrated to generate transcripts of broadcast news data. This report presents the setup and evaluation of a speech recognition system for Spanish broadcast news. Section 3 gives a short overview about the general basic components of a ASR system. Section 4 decribes the development and training process of acoustic and language models for the Spanish ASR. The performance evaluation results are dicussed in section 5. The report ends with conclusions and future work suggestions.