[Feature extraction] based on matlab formant estimation [including Matlab source code 550 period]

1. Introduction

The formant refers to some areas where the energy is relatively concentrated in the sound spectrum. The formant is not only a determinant of sound quality, but also reflects the physical characteristics of the soundtrack (resonant cavity).

The peak position on the envelope curve of the vowel and consonant sound spectrum. The original meaning of formant refers to the resonance frequency of the acoustic cavity. In the production of vowels and consonants, the sound source spectrum is modulated by the acoustic cavity. The original harmonic amplitude no longer decreases with the increase in frequency, but some strengthen and some weaken, forming a new package with ups and downs. The frequency value at the peak of the curve is consistent with the resonance frequency of the acoustic cavity. As far as vowels are concerned, the first three formants have qualitative regulations for their timbre; the first two formants are particularly sensitive to the height of the tongue, and the acoustic vowel map is drawn based on the frequency values of these two formants. of. The formant three-dimensional sonogram is shown as energy-concentrated horizontal bars.

Formant is an important feature reflecting the resonance characteristics of the vocal tract. It represents the most direct source of pronunciation information, and people use formant information in speech perception. Therefore, the formant is a very important characteristic parameter in speech signal processing, and it has been widely used as the main characteristic of speech recognition and the basic information of speech coding transmission. The formant information is contained in the frequency envelope. Therefore, the key to extracting formant parameters is to estimate the natural speech spectral envelope. Generally, the maximum value in the spectral envelope is the formant.

The principle of formant generation and its manifestation in sound quality

The distribution position of the formant is based on the resonance physical structure of the sound producing medium (Resonant Physical Structure).

Whether it is a human voice or a musical instrument, their sound characteristics are derived from two factors, one is the sound system, such as the human vocal cord or the vibrating reed of the musical instrument, and the other is the resonance system. The different resonance systems of the musical instrument make the amplitude of the partial notes in a certain frequency domain prominent. In this way, these areas produce resonance peaks unique to this musical instrument. These resonance peaks are closely related to the size and shape of the resonance body. Since the structure of a musical instrument is stable, all tones emitted by a musical instrument, regardless of the fundamental frequency, will show the same resonance peak, but its significance is strong or weak. This can help us explain why in many instruments, different tones produced by the same instrument have the same sound quality.

In speech acoustics, the human voice also has its own formant area affected by its own physiology such as the size of the nostrils, pharyngeal cavity, and oral cavity. By taking advantage of the different changes in the shape and size of these resonance spaces (such as changing the shape of the throat and mouth), we can change the formant of the sound. The reason why we can distinguish between different human voices and vowels is mainly based on the position of their formant distribution.

1 What is the role of formant and formant I have
　　talked about the excitation model of speech before, when the glottal periodic pulse excitation signal passes through the vocal tract, it will cause resonance in the vocal tract (this process is called the vocal tract model in the speech production model ), resulting in a set of resonance frequencies. This set of resonance frequencies is called formants (frequency). It is generally considered that the several maximum values in the speech spectrum envelope are formant frequencies. Accurate detection of formant frequency and bandwidth is helpful to distinguish different finals and improve the recognition of speech semantics.

2 Difficulties in the estimation of resonance peaks
(1) The existence of false resonance peaks.
(2) It is difficult to distinguish when the adjacent formant frequencies are relatively close to each other.
(3) It is difficult to extract high-pitched speech.
In short, as with pitch period estimation, there is currently no fully accurate estimation method.

3 Formant estimation pre-processing
(1) Pre-emphasis. The purpose is to remove the influence of the lip radiation and facilitate the analysis of the vocal tract response.
(2) Endpoint detection. The purpose is the same as the pitch period estimation, and there is no need to analyze the speech in the silent segment.

4 Method of resonant peak estimation
(1) Cepstrum method.
　　Cepstrum processing for speech can separate the excitation signal from the vocal tract response, then remove the excitation signal, and then do Fourier transform to get the envelope of the vocal tract response, and find the maximum value on the envelope. It is the corresponding formant frequency.
(2) In the LPC method,
　　we can obtain a set of prediction coefficients through linear prediction. According to this set of coefficients and the all-pole channel response model, perform FFT transformation to obtain the power spectrum of the channel transfer function, and then calculate the corresponding Maximum value, you can get the corresponding formant frequency.
(3) The HHT
　　method hasn’t taken a closer look, and the specific steps are not yet known. It is mainly through empirical mode decomposition (EMD) and Hilbert transform to find the amplitude of the signal , Frequency, phase.

Second, the source code

%实倒谱法共振峰估计
clear all; clc; close all;

waveFile='C4_3_y.wav';               % 设置文件名
[x, fs, nbits]=wavread(waveFile);                 % 读入一帧数据
u=filter([1 -.99],1,x);                                   % 预加重
wlen=length(u);                                          % 帧长
cepstL=6;                                                   % 倒频率上窗函数的宽度
wlen2=wlen/2;               
freq=(0:wlen2-1)*fs/wlen;                          % 计算频域的频率刻度
u2=u.*hamming(wlen);		                      % 信号加窗函数
U=fft(u2);                                                 % 按式(4-26)计算
U_abs=log(abs(U(1:wlen2)));                     % 按式(4-27)计算
 [Val,Loc,spect]=Formant_Cepst(u2,cepstL);       % 计算出共振峰频率
FRMNT=freq(Loc);                                 % 计算出共振峰频率
subplot(211)
plot(freq,U_abs,'k'); 
xlabel('频率/Hz'); ylabel('幅值/dB');
title('(a)信号对数谱X\_i(k)')
axis([0 4000 -6 2]); grid;
subplot(212)
plot(freq,spect,'k','linewidth',2); 
%LPC内插法的共振峰估计
clear all; clc; close all;

fle='C4_3_y.wav';                            % 指定文件名
[x,fs]=wavread(fle);                        % 读入一帧语音信号 
u=filter([1 -.99],1,x);                     % 预加重
wlen=length(u);                             % 帧长
p=12;                                       % LPC阶数
freq=(0:256)*fs/512;                        % 频率刻度

[F,Bw,pp,U]=Formant_Interpolation(u,p,fs);          %LPC内插法求共振峰
plot(freq,U,'k');
title('声道传递函数功率谱曲线');
xlabel('频率/Hz'); ylabel('幅值');
ll=length(F);                             % 共振峰个数
for k=1 : ll
    line([F(k) F(k)],[0 pp(k)],'color','k','linestyle','-.');    
end
% LPC求根法的共振峰估计
 clc; close all;

fle='C4_3_y.wav';                            % 指定文件名
[xx,fs]=audioread(fle);                       % 读入一帧语音信号

u=filter([1 -.99],1,xx);                    % 预加重
wlen=length(u);                             % 帧长
p=12;                                       % LPC阶数
n_frmnt=4;                                  % 取四个共振峰
freq=(0:256)*fs/512;                        % 频率刻度
df=fs/512;                                  % 频率分辨率

[F,Bw,U]=Formant_Root(u,p,fs,n_frmnt);
plot(freq,U,'k');
title('声道传递函数功率谱曲线');
xlabel('频率/Hz'); ylabel('幅值/dB');
p1=length(F);                              % 在共振峰处画线
m=floor(F/df);
pp=U(m);                                    %共振峰幅度
for k=1 : p1
    line([F(k) F(k)],[-5 pp(k)],'color','k','linestyle','-.');
end

Three, running results

Insert picture description here

Four, remarks