FPGA and configurable computing-based DSP algorithms have demonstrated significant performance improvements over software implementations. This has caused renewed interest in developing or mapping DSP algorithms to custom hardware. An algorithm will be successfully mapped if the intermediate wordlengths can be reduced to maintain reasonable hardware size. We consider linear hardware cost functions, for which we can derive closed-form expressions for the reduced wordlengths. We then apply these results to an adaptive LMS filter, where we adapt not only the tap weights, but also the wordlengths as a function of the data in real-time.

This paper explores issues arising from designing digital signal processing algorithms for dynamically-varying computing environments such as an unreliable network of processors. We present a language for specifying signal processing algorithms which permits the execution path of the algorithm to be dynamically chosen. The language leads naturally to a graphical representation of the algorithm with interesting interpretations. Finally, we formulate and characterize the solution for the problem of dynamically and optimally choosing the execution path of algorithms to minimize a system-wide cost function such as expected congestion.

In an effort to reduce the degradation in speech recognition performance caused by variation in vocal tract shape among speakers, a frequency warping approach to speaker normalization is investigated. A set of low complexity, maximum likelihood based frequency warping procedures have been applied to speaker normalization for a telephone based connected digit recognition task. This paper presents an efficient means for estimating a linear frequency warping factor and a simple mechanism for implementing frequency warping by modifying the filterbank in mel-frequency cepstrum feature analysis. An experimental study comparing these techniques to other well-known techniques for reducing variability is described. The results have shown that frequency warping is consistently able to reduce word error rate by 20% even for very short utterances.

It has been observed empirically that each coefficient in a Parks-McClellan filter converges to a steady state value as the filter length increases. This suggests the possibility of obtaining filters that are near optimal while "re-using" filter coefficients from shorter filters in the design of longer filters. In the context of approximate processing this then allows a filtering operation to be done in stages. This paper demonstrates this observation and examines some of its implications.

In an effort to reduce the degradation in speech recognition performance caused by variation in vocal tract shape among speakers, a frequency warping approach to speaker normalization is investigated. A set of low complexity, maximum likelihood based frequency warping procedures have been applied to speaker normalization for a telephone based connected digit recognition task. This paper presents an efficient means for estimating a linear frequency warping factor and a simple mechanism for implementing frequency warping by modifying the filter-bank in mel-frequency cepstrum feature analysis. An experimental study comparing these techniques to other well-known techniques for reducing variability is described. The results showed that frequency warping was consistently able to reduce word error rate by 20% even for very short utterances.