The ESTER phase II evaluation campaign for the rich transcription of French broadcast news

Abstract

This paper gives the final results of the ESTER evaluation campaign which started in 2003 and ended in January 2005. The aim of this campaign was to evaluate automatic broadcast news rich transcription systems for the French language. The evaluation tasks were divided into three main categories: orthographic transcription, event detection and tracking (e.g. speech vs. music, speaker tracking), and information extraction. The last one, limited to named entity detection in this evaluation, was a preliminary test. The paper reports on protocols and gives the results obtained in the campaign.

Constraint selection for topic-based MDI adaptation of language models

“…Our ASR system is a multipass 65K words system based on two general-purpose LMs: a 3-gram LM to create word graphs from acoustic features and a 4-gram LM to score word graphs. Experiments are carried out on 172 thematically coherent segments from 6 hours of Broadcast News (BN) shows from the French radio BN corpus ESTER [13]. These segments, coming from 3 different broadcasters and all dated from the same period of time, are spread over diversified topics (war in Iraq, national politics, sports, weather, etc.)…”

Using broad phonetic classes to guide search in automatic speech recognition

“…Experiments are conducted on radio broadcast news in French language from the ESTER2 campaign [12], which contain news shows with regular broadcast speech (RFI), but also difficult tasks like debates (Inter) Feature selection resulted in 62 attributes, containing only few MFCCs, mostly derivatives. For computational reasons, we first used the 35 best ranked APs with 10 bagging iterations and 30% sample size to optimize classification parameters like pruning strategy or splitting criterion, considering only the wide band speech during training and testing.…”

Lexical-phonetic automata for spoken utterance indexing and retrieval

“…The audio data used for the experiments consists of 6 hours of French radio broadcast news material extracted from the ESTER2 corpus [8] containing reference transcripts with manually annotated named entities. The ASR system is a large vocabulary (65k words) transcription system for which the word error rates on this corpus vary between 16.0% and 42.2%.…”