A parallel processing environment for speech signal processing applications

The paper describes a software tool that has been developed for production of speech signal processing systems, but may be easily expanded to include any 1 dimensional signal processing problem with multiple inputs and outputs. The tool allows the user to define an arbitrary set of interconnected signal processing functions that are interlinked by vectors of signal data, this kind of problem is thus expressed in an inherently MIMD format. This user defined structure is parsed to generate a configuration description that follows a communication based parallel processing format such as that used in Occam. This configuration may be executed either on a serial machine or on a network of transputers using both task and pipeline parallelism. The current system is capable of building a wide range of structures, for applications such as vocoders, speech enhancement, filterbank speech analysis and speech recognition. The parallel version runs on a network of 1825 MHz T800 transputers. Current efficiency levels are very structure dependent, but for the problems discussed within the paper, efficiencies of around 65% are achieved.