Margens de tolerância e valores de referência para os formantes de vogais orais para uso em terapias de voz para surdos em computador comercial

Abstract

RESUMO Objetivo Este trabalho apresenta as margens de tolerâncias mínimas e máximas para as frequências dos três primeiros formantes (F1, F2 e F3) na pronúncia das vogais orais do português brasileiro para utilização em terapias de voz para surdos. Método As frequências foram obtidas a partir da colaboração voluntária de 53 indivíduos adultos que tiveram as vozes gravadas e convertidas em sinais digitais durante a emissão de cada uma das sete vogais (/a/, /e/, /Ɛ/, /i/, /o/, /ᴐ/, /u/), de maneira sustentada por cerca de um segundo. As amostras foram distribuídas em dois grupos, um masculino e outro feminino. A gravação e a extração dos formantes foram efetuadas através de software próprio desenvolvido para este fim na plataforma MATLAB, utilizando o algoritmo LPC (Linear Predictive Coding) de oito coeficientes. Resultados Os resultados mostraram que uma referência consistente para os valores médios das frequências de F1, F2 e F3 pôde ser obtida através da análise gráfica e estatística das amostras de sinais de voz coletada. Conclusão Os valores de referência encontrados foram analisados e podem ser usados para calibração de dispositivos e podem servir de base para o treinamento de oralização para surdos.

Forensic Speaker Verification Using Ordinary Least Squares

“…The next step was to estimate the number of coefficients for each data windowed, aiming to extract the pitch and the formant frequencies of the signal via LPC [33]. Then, the frequencies were obtained according to the minimal frequency criteria (according to the windowing stage of the signal) and to the maximum acceptable bandwidth [26]. The frequencies that matched both criteria were considered, the first being pitch (F0) , and the subsequent being F1 to F4 .…”

“…The frequencies that matched both criteria were considered, the first being pitch (F0) , and the subsequent being F1 to F4 . In the literature, it was demonstrated that these are liable to encompass the spectral information regarding each vowel [26]. However, the developed software runs on extracting the greatest possible number of formants from audio (approximately 16–20 formants).…”

“…For audio analysis, spectral analysis is a valuable tool, because it extracts information about the origin and properties of the audio file, assessing the most prominent frequency components of the signal. These most prominent frequencies are known as formants and they work as a species of a particular signature for each presenter, providing the intrinsic properties of the vocal tract for each person [26].…”

“…They are characterised by peaks that appear throughout the spectrum. A number of studies have considered the major contribution of up to five of those frequencies on voice composition, the first being of greater intensity and denominated pitch (or F0 ), and the others ( F1 to F4 ) accordingly [26].…”

“…With the roots obtained by the LPC model, it is possible to calculate the formant frequencies, using the tangent arch coordinated on the unitary circle [35,36,37], which reflect the peaks estimated by the LPC model [33]. The criteria commonly used for formant frequencies selection by LPC considers 90 Hz as a minimum frequency, and the bandwidth is less than 400 Hz [26]. Figure 1 presents the spectrum obtained by the decomposition of a given audio signal on its frequency components, and the envelope through LPC.…”