Distributed speech recognition in the presence of noise and packet loss

Abstract

This paper investigates the benefits of combining speech enhancement and packet loss mitigation in a Distributed Speech Recognition (DSR) system operating over a packet-based network. The recognition task is connected digits with feature vectors extracted from speech using two different auditory-based front-ends. Feature vectors are vector quantized and combined in pairs for transmission over a channel that is subject to packet loss. Results demonstrate that the combination of speech enhancement and packet loss mitigation techniques can greatly improve the speech recognition performance of a DSR system.The use of voice as a user interface technology for data entry to network services is increasing. For mobile wireless devices, small keypads and displays can be an inconvenient method of data entry and the use of reliable voice input can offer a greatly improved input method. With the increasing use of the Internet for the transmission of voice over IP (VoIP), new voice-enabled data services and applications are being developed by service providers. To facilitate access to services by means of voice input there is a need to include automatic speech recognition (ASR) as part of the user interface.In Distributed Speech Recognition (DSR), the speech recognition task is split between the terminal, where the front-end feature extraction is performed, and the network, where the back-end recognition is performed. The features that represent the speech are sent by means of an error protected data channel to the classifier for processing. DSR avoids both the speech encoding and decoding stages associated with centralized (server-based) ASR systems and so eliminates the degradations that originate from the speech compression algorithms that are used in order to achieve low bit-rate transmission. The European Telecommunications Standards Institute (ETSI) published their first standard for DSR in 2000 [1] under the Aurora working group.Recent comparative studies have shown the superior performance of DSR to codec-based ASR [2]. However, in a DSR system transmission errors in the form of random packet loss and packet burst loss still need to be taken into consideration. Such transmission errors can have a significant impact on recognition accuracy. Furthermore, it is well known that the presence of noise severely degrades the performance of speech recognition systems. A large amount of research has been carried out on speech enhancement, and a number of approaches have been well documented in the literature, for example [3]. ETSI has published a second DSR standard [4] which includes enhancement of the speech signal prior to feature extraction.In this paper, the speech recognition performance of the ETSI advanced front-end for DSR [4] is compared with the auditory front-end proposed by Li et al. [5], in the presence of the combined degradations of background noise and packet loss. Speech enhancement and error mitigation are used to compensate for noise and packet loss. The recognition problem examined is connected digit recogn...