使用matlab基于卷积神经网络实现的汉语短句语音识别，输出拼音.zip

function [Weight,CorrectRate] = CalcCTC(NN,NNInput,charFeature,usedNNOutputs) NNOutput = NN.forward(NNInput); NNOutput(:,:,1:2 ,:) = softmax(NNOutput(:,:,1:2 ,:));% 1>2表示有声 否则表示没声（间断） NNOutput(:,:,3:26 ,:) = softmax(NNOutput(:,:,3:26 ,:));% 声母 NNOutput(:,:,27:61,:) = softmax(NNOutput(:,:,27:61,:));% 韵母 NNOutput(:,:,62:66,:) = softmax(NNOutput(:,:,62:66,:));% 音调 % 概率计算负对数，方便后面用加法进行条件概率计算 epsilon=1e-5; NNOutput(NNOutput < epsilon)=NNOutput(NNOutput < epsilon)+epsilon; % CTC一堆标量操作 用gpu算更慢 这里把数据取到CPU算 NNOutput=gather(extractdata(-log(NNOutput))); Weight = zeros(size(NNOutput)); Loss = 0; for i=1:size(NNInput,4) % 计算CTC得到序列映射关系矩阵 [loss,Weight(:,1:usedNNOutputs(i),:,i)] = CTCLoss(NNOutput(:,1:usedNNOutputs(i),:,i), charFeature{i}); Loss=Loss+loss; end % 统计正确率 % softmax输出转换成编号 % 前面取了负对数所以这里要最小值 NNCategorical = zeros([size(NNOutput,[1 2]) ,4,size(NNOutput,4)] ); [~,NNCategorical(:,:,1,:)] = min(NNOutput(:,:,1:2,:),[],3); [~,NNCategorical(:,:,2,:)] = min(NNOutput(:,:,3:26,:),[],3); [~,NNCategorical(:,:,3,:)] = min(NNOutput(:,:,27:61,:),[],3); [~,NNCategorical(:,:,3,:)] = min(NNOutput(:,:,62:66,:),[],3); % CTC得到的匹配矩阵转换成编号 WeightCategorical = zeros([size(Weight,[1 2]) ,4,size(Weight,4)] ); [~,WeightCategorical(:,:,1,:)] = max(Weight(:,:,1:2,:),[],3); [~,WeightCategorical(:,:,2,:)] = max(Weight(:,:,3:26,:),[],3); [~,WeightCategorical(:,:,3,:)] = max(Weight(:,:,27:61,:),[],3); [~,WeightCategorical(:,:,3,:)] = max(Weight(:,:,62:66,:),[],3); num=0;denum=0; for i=1:size(Weight,4) % 识别出来是静音（WeightCategorical（:,:,1,:)==2)的段不参与统计 NNCategorical_Used = WeightCategorical(1,1:usedNNOutputs(i),1,i) == 1; % 多少个字识别对了(有无，声母，韵母，音调都识别对） num = num+sum(all(NNCategorical(1,NNCategorical_Used,:,i) == WeightCategorical(1,NNCategorical_Used,:,i),3),'all'); % 一共有多少个字 denum = denum+sum(NNCategorical_Used); end CorrectRate = 100*num/denum; end