Infoscience

Report

Adaptation Experiments on French MediaParl ASR

This document summarizes adaptation experiments done on French MediaParl corpus and other French corpora. Baseline adaptation techniques are briefly presented and evaluated in the MediaParl task for speaker adaptation, speaker adaptive training, database combination and environmental adaptation. Results show that by applying baseline adaptation techniques, a relative WER reduction of up to 22.8% can be reached in French transcription accuracy. For the MediaParl task, performance of systems trained on directly merged databases and of systems trained on databases combined via MAP adaptation did not differ significantly when large amount of data was available. During the experiments, French data recorded in Switzerland behaved in a similar way compared to French data recorded in France, which suggest that French spoken in Valais is close to the standard French spoken in France, and differencies in ASR accuracies between models trained on Swiss MediaParl and on French BREF are more likely caused by environmental factors or more spontaneity in speech.

    Reference

    • EPFL-REPORT-192654

    Record created on 2013-12-19, modified on 2016-08-09

Related material