Using smoothed K-TSS language models in continuous speech recognition

k-TSS language models in speech recognition systems

“…Thus, many authors share the objective of obtaining accurate language models that can be well integrated into the maximization procedure in Equation (2), and can be handled in a one-step decoding procedure (Placeway, Schwartz, Fung & Nguyen, 1993;Riccardi et al, 1996;Bonafonte & Mariño, 1998;Llorens, 2000). This is also one of the main goals of our work (Bordel, Torres & Vidal, 1994;Varona & Torres, 1999;Torres & Varona, 2000;Varona, 2000).…”

“…This formulation leads to a very compact representation of the model parameters learned at training time, i.e. probability distribution and model structure (Varona & Torres, 1999;Torres & Varona, 2000) (see Section 5). This compact structure allows non-deterministic parsing when used in a CSR system leading to different probability distributions.…”

“…Its aim is to show the ability of stochastic regular grammars to generate accurate language models which can be well integrated into the full scheme given by Equation (2). For this purpose, a stochastic grammar generating a certain subclass of regular languages called k-testable language in the strict sense (k-TSS) (García & Vidal, 1990;Segarra, 1993;Bordel et al, 1994;Varona & Torres, 1999, 2000bTorres & Varona, 2000) is considered. In previous works (Bordel et al, 1994;Varona & Torres, 1999) the use of k-TSS language models has been proposed for integration into a CSR system.…”

“…For this purpose, a stochastic grammar generating a certain subclass of regular languages called k-testable language in the strict sense (k-TSS) (García & Vidal, 1990;Segarra, 1993;Bordel et al, 1994;Varona & Torres, 1999, 2000bTorres & Varona, 2000) is considered. In previous works (Bordel et al, 1994;Varona & Torres, 1999) the use of k-TSS language models has been proposed for integration into a CSR system. Given a positive sample R + of strings of an unknown language, the inference algorithm to obtain a deterministic finite-state automaton that recognizes the smallest k-TSS language containing R + is proposed in García and Vidal (1990).…”

“…(2) Direct representation of the whole network representing the SFSA is usually prohibitive even for small vocabulary tasks (Riccardi et al, 1996;Mohri et al, 2000;Varona & Torres, 2000b). (3) Smoothing procedures need to be developed under the syntactic approach (Bordel et al, 1994;Riccardi et al, 1996;Varona & Torres, 1999, 2000aLlorens, 2000). The recursive schema usually involved in the smoothing procedure must be adopted in the finite state formalism to achieve efficient implementations while obtaining adequate probability distributions.…”