The covariance matrix in GMM go to singular. How can I fix it ? and Where is the problem ?

I am implementing GMM in matlab. But when I run that code, the covariance go to singular after 2 interations in EM-Algorithm. I don't know where is the problem. Data is 143-dimensional vector and about 300 observers. This is my code:

for (iter = 1:100)

    fprintf('  EM Iteration %d\n', iter);

    %%==========================a====================
    %%STEP 3a: Expectation
    % tinh lai trong so
    pdf = zeros(m, k);
    % For each cluster...
    for (j = 1 : k)

        % Tinh xac suat Gauss cho tat ca cac diem thuoc cluster 'j'.
        pdf(:, j) = gaussianND(X, mu(j, :), sigma{j});  %<======this is my problem. Det(sigmal{j}) =0;
    end

    %    pdf  [m  x  k]
    %    phi  [1  x  k]   
    %  pdf_w  [m  x  k]
    pdf_w = bsxfun(@times, pdf, phi);

    % Divide the weighted probabilities by the sum of weighted probabilities for each cluster.
    %   sum(pdf_w, 2) -- sum over the clusters.
    W = bsxfun(@rdivide, pdf_w, sum(pdf_w, 2));

    %%===============================================
    %%STEP 3b: Maximization
    %%Calculate the probability for each data point for each distribution.

    % store mean
    prevMu = mu;    

    % Cap nhat cho moi cluster
    for (j = 1 : k)

        % update weight
        phi(j) = mean(W(:, j), 1);

        % update mean.
        mu(j, :) = weightedAverage(W(:, j), X);

        % update covariance

        sigma_k = zeros(n, n);
        Xm = bsxfun(@minus, X, mu(j, :));
        for (i = 1 : m)
            sigma_k = sigma_k + (W(i, j) .* (Xm(i, :)' * Xm(i, :)));
        end

        % Divide by the sum of weights.
        sigma{j} = sigma_k ./ sum(W(:, j));
    end

    % Kiem tra hoi tu
    if (mu == prevMu)
        break
    end         
end

Anyone has answer?