Abstract: This paper reports our efforts on the adaptation of a baseline system trained on clean speech to a task for which French native speakers uttered some Spontaneous French queries while driving a car. When the system is retrained on the new task acoustic data the Word Error Rate (WER) is decreased by 60% compared to our baseline system initial performance on the new task. We show that on spontaneous queries, L of this improvement could be achieved without prior system retraining by a more accurate Language Modelling which takes into account the noises and spontaneous speech effects and by a carefull grapheme/phoneme transcription of foreign words. We also describe the integration of this French system in our Multilingual Navigation System.
Loading