An algorithm for fast composition of weighted finite-state transducers

Abstract

In automatic speech recognition based on weighted-finite transducers, a static decoding graph HC • L • G is typically constructed. In this work, we first show how the size of the decoding graph can be reduced and the necessity of determinizing it can be eliminated by removing the ambiguity associated with transitions to the backoff state or states in G. We then show how the static construction can be avoided entirely by performing fast on-the-fly composition of HC and L • G. We demonstrate that speech recognition based on this on-the-fly composition approximately 80% more run-time than recognition based on the statically-expanded network R, which makes it competitive compared with other dynamic expansion algorithms that have appeared in the literature. Moreover, the dynamic algorithm requires a factor of approximately seven less main memory as the recognition based on the static decoding graph.

Use of contexts in language model interpolation and adaptation

“…Component n-gram probabilities and interpolation weights will be applied for each context on request during decoding. Similar approaches have been previously shown to be effective for the composition between one single back-off n-gram LM and a lexicon transducer (Caseiro and Trancoso, 2006;Cheng et al, 2007;McDonough et al, 2007;Oonishi et al, 2009).…”

Accurate client-server based speech recognition keeping personal data on the client

“…Finally, M 1 is composed on-the-fly with a hidden Markov model H along with the cross-word computation. This was initially described by Hori et al [22], [23] and further improved by McDonough et al [24] and Allaucen et al [25], [26]. The probability for a phoneme sequence given a sequence of speech features can be computed [27] using a token passing time synchronous Viterbi beam search [28].…”

A Filter-Based Algorithm for Efficient Composition of Finite-State Transducers

“…If the new filter state is not the blocking state ⊥ and a new transition is created from the filter-rewritten transitions (e ′ 1 , e ′ 2 ) (line 14). If the destination state (n[e ′ 1 ], n[e ′ 2 ], q ′ 3 ) has not been found previously, it is added to Q and inserted in S (lines [11][12][13]. The composition algorithm presented here is available in the OpenFst library [3].…”

“…For some problems, it is possible to find equivalent inputs that will compose more efficiently, but it is not always possible or desirable to do so. This has been especially an issue in natural language processing applications and led to special-purpose composition algorithms for use in speech recognition [6,7,11,15] and speech synthesis [2].…”