Audio packet loss over IP and speech recognition

This paper deals with the effects of packet loss on speech recognition over IP connections. The performance of our continuous French speech recognition system is here evaluated for different transmission scenarios. A packet loss simulation model is first proposed in order to simulate different channel degradation conditions. The packet loss problem is also investigated in real transmissions through IP. Because packet loss impact may be different according to the speech coder used to transmit data, different transmission conditions with different audio codecs are also investigated. Several reconstruction strategies to recover lost information are then proposed, and tested. Another solution for dialog applications is also suggested, where the relative weight of the language and acoustic model is changed according to the packet loss rate. The results show that the speech recognition performance can be augmented by the solutions here presented.