Acquisition of an extensive rule set for slovene grapheme-to-allophone transcription

Abstract

This paper describes a prospective study of the contribution of a single-sensor noise pre-processing method, prior to coding, to the performance of a parametric low bit rate speech coder in adverse conditions. The 2.4kbits/s vocoder we use estimates four parameters: fundamental frequency, voicing, linear prediction coefficients and energy. Firstly, we study the influence of different noise levels on the estimated parameters with and without noise reduction system. Secondly, we measure the contribution of (i) each speech coder parameter and (ii) the speech enhancement system to the global output intelligibility. Finally, results show the interest of such a speech enhancement system for low bit rate parameter estimation and underline the interest to adapt different pre-processing techniques for each parameter estimation.

Spectral enhancement preprocessing for the HNM coding of noisy speech

“…Reference parameters obtained from the HNM analysis of the clean speech are used to determine parameter errors that occur after the addition of noise. A pitch estimate is deemed correct if its value does not differ from the reference by more than 5% [1].…”

“…Recently, significant improvements have been reported when specific speech coders were combined with a speech enhancement preprocessor. Guilmin et al [1] showed that Wiener filter-based noise preprocessing significantly improved the output, in the presence of noise, of a low rate LPC vocoder both in terms of parameter estimates and subjective quality. Earlier Kang and Fransen [2] evaluated spectral subtraction enhancement for LPC-processing of noisy speech and reported dramatic improvements in subjective quality for speech corrupted with a variety of background noise.…”