Parallel Structure in an Integrated Speech-Recognition Network

Abstract

Abstract. Large-vocabulary continuous-speech recognition (LVCR) speakerindependent systems which i n tegrate cross-word context dependent acoustic models and n-gram language models are di cult to parallelize because of their interwoven structure, large dynamic data structures, and complex object-oriented software design. This paper shows how retrospective decomposition can be achieved if a quantitative analysis is made of dynamic system behaviour. A design which accommodates unforeseen e ects and future modi cations is presented.

Performance metrics for embedded parallel pipelines

“…p i ) + 1 2 G 00 (U i:p ? p i ) 2 : : : ; (21) where G 0 denotes d=duG(u)j u=p i ; u = F(x). A convenient example is a logistics cdf, F(y) = Figure 16 uses expansion (22) to make a continuous estimate of E Y i:p ], i odd.…”

“…However, irregular algorithms which involve global access to a common data structure are not suitable for PPF as the essence of the data farm paradigm is strictly local computation. For example, compare the parallel pipeline for speech processing reported in 3] in which the elements of a decoding network have been decomposed and are suitable for PPF development to the integrated speech decoder network reported in 21], which was unsuitable for the PPF approach.…”

Semi-structured portable library for multiprocessor servers

“…From these frames, a speech recognition algorithm, such as one based on Hidden Markov Models (HMM's), extract words and sentences. It has already been established [5] that speech recognition involves both computationally intensive processing to form the Gaussian mixture weightings at each node of the HMM, and synchronized access to a complex data structure to update the sentence model as more frames arrive. Subsequently, identified text has to be interpreted to extract its meaning.…”