www.gusucode.com > 端点检测和基于DTW和HMM的孤立词识别和连续语音识别 > code/cdhmm/train.m

    function [hmm, pout] = train(samples, M)
%输入:
%  samples -- 样本结构
%  M       -- 为每个状态指定pdf个数,如:[3 3 3 3]
%输出:
%  hmm      -- 训练完成后的hmm

K   = length(samples);

% 计算语音参数
disp('正在计算语音参数');
for k = 1:K
	if isfield(samples(k),'data') & ~isempty(samples(k).data)
		continue;
	else
		samples(k).data = mfcc(samples(k).wave);
	end
end

hmm = inithmm(samples, M);

for loop = 1:40
	fprintf('\n第%d遍训练\n\n',loop)
	hmm = baum(hmm, samples);

	%计算总输出概率
	pout(loop)=0;
	for k = 1:K
		pout(loop) = pout(loop) + viterbi(hmm, samples(k).data);
	end

	fprintf('总和输出概率(log)=%d\n', pout(loop))

	%比较两个HMM的距离
	if loop>1
		if abs((pout(loop)-pout(loop-1))/pout(loop)) < 5e-6
			fprintf('收敛!\n');
			return
		end
	end
end

disp('迭代40次仍不收敛, 退出');