Fernando Gil Resende

A rule-based grapheme-phone converter and stress determination for Brazilian Portuguese natural language processing

This paper presents a grapheme-phone converter and stress determination algorithm based on rules. The proposed set of rules was implemented and tested on a randomly chosen extract of the CETEN-Folha text database. Computer experiments show it is achieve an accuracy of 97.44% and 98.58%, respectively, for the grapheme-phone converter and the stress determination algorithm.

Grapheme-phone transcription algorithm for a Brazilian Portuguese TTS

This paper describes one of the aspects of an ongoing research to improve the synthetic speech quality for a Brazilian Portuguese text-to-speech synthesis system. This paper focuses on the grapheme-phone transcription algorithm. A complete set of rules for every grapheme is presented. Experimental results, obtained running the proposed algorithm through a text database, gave rise to 98,43% of accuracy rate.

New adaptive algorithms based on multi-band decomposition of the error signal

New adaptive algorithms based on multi-band decomposition of the error signal and application of a different convergence factor for each band are derived. With this approach, tracking ability and performance in steady state can be traded off along the frequency domain giving rise to estimates of the adaptive filter coefficients closer to the ideal response as compared to those obtained with conventional least-mean-square (LMS) and recursive least-squares (RLS) algorithms, particularly when the statistical properties of the analyzed signal vary along the frequency spectrum. A new adaptation technique for the forgetting factor depending exclusively on the autocorrelation values of the input signal is also introduced and a multi-band RLS algorithm, with an independent variable forgetting factor for each band, suitable for the analysis of nonstationary signals is described. Computer experiments comparing the performance of multi-band and conventional LMS and RLS algorithms are shown.

AR spectrum estimation based on wavelet representation

A new adaptive AR spectrum estimation method is proposed. The cost function is defined by using the discrete-time wavelet transform coefficients of the linear prediction error. Instead of a single window throughout the whole frequency spectrum, a wavelet-like windowing method is used to increase the frequency resolution of the low-frequency components and to improve the time resolution of the high-frequency components. Special properties of the covariance matrix are used to derive an RLS algorithm which requires O(M/sup 2/) operations. Simulation results show that the wavelet based spectrum estimation method gives fine frequency resolution at low frequencies and good time resolution at high frequencies, while with conventional methods it is possible to have only one of these characteristics.<<ETX>>

AR spectral estimation based on multi-window analysis of the linear prediction error

A new method for autoregressive (AR) spectral analysis and a fast-transversal-filters (FTF) recursive algorithm are introduced. While conventional least-squares (LS) methods use a single windowing function in the analysis of the linear prediction error, the proposed method decomposes the linear prediction error into several bands and analyzes each of them through a different window. With this approach, the variance of spectral estimates and the tracking ability of the spectral analyzer can be traded off throughout the frequency spectrum, giving rise to spectral estimates that represent the true underlying spectrum with better fidelity than conventional LS methods. Mathematical background for the design of fast recursive algorithms for multi-window LS is exposed and an FTF algorithm is derived. Simulations comparing the performance of conventional and multi-window LS are shown.

Sparse KPCA for feature extraction in speech recognition

This paper presents an analysis of the applicability of sparse kernel principal component analysis (SKPCA) for feature extraction in speech recognition, as well as a proposed approach to make the SKPCA technique realizable for a large amount of training data, which is a usual context in speech recognition systems. Although the KPCA (kernel principal component analysis) has proved to be an efficient technique for being applied to speech recognition, it has the disadvantage of requiring training data reduction, when its amount is excessively large. The standard approach to perform this data reduction is to randomly choose frames from the original data set, which does not necessarily provide a good statistical representation of the original data set. In order to solve this problem a likelihood related re-estimation procedure was applied to the KPCA framework, thus creating the SKPCA. The experimental results show the efficiency of SKPCA technique with the proposed approach over the KPCA with the standard sparse solution using randomly chosen frames and the standard feature extraction techniques.

Mixed-excited phonetic vocoding at 265 bps

In this paper a phonetic vocoder which synthesizes speech using mixed excitation is presented. The encoder carries out HMM-based speech recognition and pitch analysis, whereas the decoder performs parameter extraction from HMM and builds a mixed excitation using pitch and bandpass voicing strengths. The vocoder at an average bit rate of 265 bit/s reaches good degree of intelligibility, while the use of mixed excitation significantly improves the speech quality with no increase of bit rate when compared with the conventional binary excitation pulse train/random noise.

Adaptive AR spectral estimation based on multi-band decomposition of the linear prediction error with variable forgetting factors

A new method for adaptive autoregressive spectral stimulation based on the least-squares criterion with multi-band decomposition of the linear prediction error and analysis of each band through independent variable forgetting factors is presented. The proposed method localizes the forgetting factor adaptation scheme in the frequency domain and in the time domain, in the sense that variations on the statistics of the input signal are independently evaluated for each band along the time. In this paper, a new forgetting factor adaptation technique depending exclusively on the input signal is introduced and applied to the multi-window analysis of the linear prediction error structure to generate time-varying autoregressive spectral estimates. An improvement on the fidelity of estimates is shown in computer experiments which compare the proposed method with conventional and multi-band least-squares methods with fixed forgetting factors.

A methodology to analyze homographs for a Brazilian Portuguese TTS system

In this work, a methodology to analyze words that are homographs and heterophones is proposed to be applied in a Brazilian Portuguese text-to-speech system. The reasoning is based on grammatical construction. An algorithm structured on the presented methodology was implemented to solve the reading decision problem for the word sede giving rise to 95,0% of accuracy rate when tested on the CETEN-Folha text database.

An on‐line laboratory course on speech analysis

An on‐line laboratory course for speech signal processing is described. The course consists of a series of practical experiments involving the three aspects of speech processing: coding, synthesis, and recognition. The experiments were developed using an auxiliary tool implemented in Delphi and were then transformed into individual class‐modules using the eTEAM software. Both pieces of software are briefly described, along with three experiments already available through the Internet.