基于HMM的非特定人孤立词语音识别MATLAB代码

%本程序应用多窗谱法估计的语音信号功率谱密度（PSD）来进行谱减语音增强 clear; a=2; %过减因子 b=0.01; %增益补偿因子 c=0; %c=0时，不对增益矩阵进行开方，c=1时，进行开方运算 %读取语音文件--------------------------------------------------------------- [filename,pathname]=uigetfile('SNR_0-增大.wav','请选择语音文件：'); [wavin_t,fs,nbits]=wavread([pathname filename]); wav_length=length(wavin_t); %基音周期最大为20ms，为使ifft还原后语音失真尽量小，帧长至少要为基音周期的2倍 %根据fs选择帧长： switch fs case 8000 frame_len=320;step_len=160; case 10000 frame_len=400;step_len=200; case 12000 frame_len=480;step_len=240; case 16000 frame_len=640;step_len=320; case 44100 frame_len=1800;step_len=900; otherwise frame_len=1800;step_len=900; end; frame_len=320;step_len=160; frame_num=ceil((wav_length-step_len)/step_len); wavin=zeros(1,frame_num*frame_len); wavin(1:wav_length)=wavin_t(:); inframe=zeros(frame_len,frame_num); for i=1:frame_num; inframe(:,i)=wavin(((i-1)*step_len+1):((i-1)*step_len+frame_len)); end; %inframe=(ENFRAME(wavin,frame_len,step_len))'; %分帧 %frame_num=size(inframe,2); %求帧数 window=hamming(frame_len); %定义汉明窗 %分别对每帧fft，求幅值，求相角----------------------------------------------- for i=1:frame_num; fft_frame(:,i)=fft(window.*inframe(:,i)); abs_frame(:,i)=abs(fft_frame(:,i)); ang_frame(:,i)=angle(fft_frame(:,i)); end; %每相邻三帧平滑------------------------------------------------------------- abs_frame_f=abs_frame; for i=2:frame_num-1; abs_frame_f(:,i)=mean(abs_frame(:,(i-1):(i+1)),2); end; abs_frame=abs_frame_f; %求增益矩阵----------------------------------------------------------------- %矩阵中每一元素为： %g(k)=(Py(k)-a*Pn(k))/Py(k) %Py和Pn分别为带噪语音和噪声的功率谱估计，都用MATLAB中自带的pmtm函数来估计 %可根据需要调节a的大小，来得到更好的效果 %用多窗谱法法对每一帧数据进行功率谱估计 for i=1:frame_num; per_PSD(:,i)=pmtm(inframe(:,i),3,frame_len,'twosided'); end; %对功率谱的每相邻三帧进行平滑 per_PSD_f=per_PSD; for i=2:frame_num-1; per_PSD_f(:,i)=mean(per_PSD(:,(i-1):(i+1)),2); end; per_PSD=per_PSD_f; %取前20帧作为噪声帧，取其平均作为噪声的功率谱估计 noise_PSD=mean(per_PSD(:,1:20),2); %求增益矩阵 for k=1:frame_num; g(:,k)=(per_PSD(:,k)-a*noise_PSD)./per_PSD(:,k); end; %求增益补偿阈值，凡是小于该阈值的增益系数军用阈值来代替，这样可减少音乐噪声 spec_floor=b*noise_PSD./per_PSD(:,k); spec_floor=spec_floor(:,ones(1,frame_num)); [I,J]=find(g<spec_floor); gf=g; gf(sub2ind(size(gf),I,J))=spec_floor(sub2ind(size(gf),I,J)); if c==0; g=gf; else g=gf.^0.5; end; %谱减---------------------------------------------------------------------- sub_frame=g.*abs_frame; % % 非语音帧衰减------------------------------------------------------------ % T=20*log10(mean(sub_frame./(abs_noise*ones(1,frame_num)))); % T_noise=mean(T(:,1:20),2); % c=10^(-2/3); %衰减系数为 10^(-1.5) % noise_frame=find(T<T_noise); % sub_frame(:,noise_frame)=c*sub_frame(:,noise_frame); %将语音信号还原至时域------------------------------------------------------- wavout=zeros(1,(frame_num-1)*step_len+frame_len); j=sqrt(-1); i=1; for t=1:step_len:((frame_num-1)*step_len+1); wavout(:,t:(t+frame_len-1))=wavout(:,t:(t+frame_len-1))+real(ifft(sub_frame(:,i).*exp(j*ang_frame(:,i))))'; i=i+1; end; %将处理结果输出为'wav'文件-------------------------------------------------- if c==0; wavwrite(wavout,fs,nbits,[num2str(frame_len) 'fnm_' num2str(a) '_' num2str(b) '_' filename]); else wavwrite(wavout,fs,nbits,['m_' num2str(a) '_' num2str(b) '_' filename]); end; %将处理前后的结果进行作图比较 subplot(2,1,1); plot(wavin);grid on; axis([1 wav_length -1 1]); subplot(2,1,2); plot(wavout);grid on; axis([1 wav_length -1 1]); sound(wavin); sound(wavout);