Indexado en

Licencia y uso

Citaciones

Altmetrics

Análisis de autorías institucional

Compartir

Lattice-based time-delay neural network for speech processing

Publicado en:Lecture Notes In Computer Science. 930 963-970 - 1995-01-01 930(), DOI:

Autores: Gomez, P; Rodellar, V; Nieto, V; Hombrados, MA

Afiliaciones

Resumen

The use of Time-Delay Neural Networks (TDNN's) in Continuous Speech Recognition has not been as relevant as it was expected due to the computational costs implied by Time-Delay orders, as it was taken for granted that the bigger the orders, the better the representation of the dynamic essence of Speech. This paper focuses on the true differential nature of this representation, and proposes to see TDNN's as devices working on differential relations among delayed versions of Speech Spectra, using Lattice Predictors as processing delay lines, which de-correlate the information which is presented to the computing nodes. This results in optimally compact structures (minimum number of delays), and better convergence rates. Convergence experiments show that reductions in the global computational costs as low as 1:5 may be achieved using structures based on this method as compared with traditional TDMN's.

Palabras clave