$%675$&7 ,1752'8&7,21

To be able to build acoustic models for children, that can be used in spoken dialogue systems, speech data has to be collected. Commercial recognizers available for Swedish are trained on adult speech, which makes them less suitable for children's computer-directed speech. This paper describes some experiments with on-the-fly voice transformation of children's speech. Two transformation methods were tested, one inspired by the Phase Vocoder algorithm and another by the Time-Domain Pitch-Synchronous Overlap-Add (TD-PSOLA) algorithm. The speech signal is transformed before being sent to the speech recognizer for adult speech. Our results show that this method reduces the error rates in the order of thirty to forty-five percent for children users. Commercially available speech recognizers of today are mostly trained on adult speech, making them less suitable for spoken input from children. Thus, when designing children-directed spoken dialogue applications, it is important to collect data under realistic circumstances in order to construct language models that work well for children. However, simple data collection methods that are used to collect adult speech, like reading prepared sentences from a paper, or having people talk to a simple system-directed dialogue system, do not work that well for children. It is problematic to record children reading aloud [1], and having children talk to a dialogue system fails because of the very problem mentioned in the beginning-the recognition performance is bad for children, making it hard to get a dialogue going. This is a bootstrapping problem-how do you collect spoken dialogue data from children (in order to improve recognition for children) without having a good recognizer available? In previous studies this problem has been solved using Wizard-of-Oz techniques to collect dialogue data from children [2,6]. However, if the dialogue system is to be used in a permanent public exhibition this solution is too expensive and complicated.Ã ÃThis article describes a method that aims at solving this problem for a dialogue system placed in a permanent exhibition. The idea is to transform the children's speech to lower frequencies before down-sampling it to telephone bandwidth. This method has been implemented in a simple system directed dialogue system that was designed to be engaging for children. It is a small speech enabled computer game publicly available in a futuristic apartment that is on display at the Telecommunication museum in Stockholm. The users interact with an animated agent to fix things in the apartment or to get to know …