Abstract
This pnper describes n sqjhunre tool thnt hns heendeveloped jiw production o f speech signnlprocessing system, but iiiqy he ensi!y expnndetl toinclude any I -ilinrensionnl signnl pi.ocessiiigprobleni with multiple inputs nnd outputs. The toolnllows the user to tlejine nn nrhitrnry set ofinterconnected signal processing junctions that oreinterlinked h,v vectors of signnl dntn, this kind ofprohlenr is thus expressed iii nn inhereentfy MIMD$ornint. This user-defined structure is pnrsed togenernte o con$gumtion description that J~llow~nsconinii~iiicatioii based pnrnllel prwcessing forinntsuch ns thnt used in Occam. This configuration ninvhe executed either on n serial niochine or on nnetwork of trnnsputers using both task and pipelinepal-allelisi~. Tlie current system is cnpable ofbuilding n wide rnnge of structures, for npplicntionssuch ns voco(Jers, speech enhnncenrent, jilterhnnkspeech nnnivsis and speech recognition. Thepnrnllel version IWIS on n network of 18 25 AdHzT800 trnnsputers. Current eflciency Iesels are verystructure &pentlent, hut jilr the prohlenw discussedwithin the paper, efficieiicies oj around 6-5 % nileachieved
.1. Introduction
This paper describes work being carried out forspeech signal processing applications. It reports onnovel parallel environment where signal processing'blocks' can be defined by the user (includingspecific parameters; such as size of FFT. differentwindowing types, and nuniber of layers in a MultilayerPerceptron). These signal processing blocksare then placed onto a large parallel processingnetwork (up to 66 T800 transputers) for efficientimplementation. This paper describes the1066-6192196 $5.00 0 1996 IEEEProceedings of PDP '96environment designed to enable the user to build upthese blocks for the processing network.It is well known that many signal processingapplications are well suited for efficientimplementation on parallel machines [6] [7].However, many of these must be 'hand-crafted' ontothese machines to achieve the best performance.What is required for signal processing designer is asystem that takes some, or all, of this hand-craftingaway, and is performed automatically by the system.Such a system is being designed and is detailed inthis paper.The current system has been tailored to be easy touse in any speech processing environment, withemphasis on speech detection, recognition andenhancement. In such areas it may be necessary toexperiment with many different signal processingstructures. A special Signal Processing DescriptorFile (SPDF) has been designed, which contains theuser information for a particular signal processingstructure. Such a file contains information about theprocessing elements that have been defined: theirfunction, the input details, their output details, andany compile-time parameters required. These filesare parsed by a specially designed parser thatproduces a configuration description file which isloaded onto the various processors in the network.A number of standard signal processing hnctionshave been 'hard-wired' into the system to allowlarger systems to be built. These include 'standard'signal processing functions such as FFTs, filters,LPC, wh~dowing functions and less 'standard'functions such as multi-layer perceptrons (MLPs) orcepstral transformations. These allow complexsystems to be built easily out of simply-definedbuilding blocks. These functional building blockshave a standard C format and new functions areeasily added to the compiled code.


Download full report
http://doi.ieeecomputersociety10.1109/EMPDP.1996.500621