Question

我尝试使用隐马尔可夫模型实现说话人验证。我遵循了本教程：http://compbio.ucdenver.edu/hunter/cpbs7711/2010_09_09_Rabiner1989_HMMTutorial_Erratum_Leach.pdf

显然，我使用转换矩阵值得到了错误的结果。我想，或者我错了。转换矩阵中每行的概率总和小于1.这应该发生在多个观察中吗？此外，观察结果应该具有相同的长度吗？我使用MFCC并使用Kmeans量化它们。我应该确保它们长度相同吗？任何帮助，将不胜感激。试图寻找其他来源，但有点丢失。

# Function for getting the xi and epsilon values
def ep_xi(self, states, obs, alphabet, init_pi, transition_matrix, emission_matrix, a, b):

ip = deepcopy(init_pi)
tm = deepcopy(transition_matrix)
em = deepcopy(emission_matrix)
coef = deepcopy(a["coef"])
alpha = deepcopy(a["alpha"])
beta = deepcopy(b)
T = len(obs)
epsilon = {}
xi = {}

epsilon[0] = None
    for t in range(1, T):

        epsilon[t] = {}

        for i in states:

            epsilon[t][i] = []


            for j in states: 
                below = 0
                epsilon_t = alpha[t-1][i] * tm[i][j] * em[j][obs[t]] * beta[t][j]
                for k in states:
                    below += (alpha[t][k] * beta[t][k])


                if below > 0: epsilon[t][i].append(epsilon_t/below)
                else: epsilon[t][i].append(0)
for t in range(0, T):

        xi[t] = []

        for i in states:
            xi_t = alpha[t][i] * beta[t][i]
            below = 0

            for j in states:
                below += (alpha[t][j] * beta[t][j])

            if below == 0: xi_t = 0 
            else: xi_t /= below 

            xi[t].append(xi_t)

# Function for reestimating the matrices values
def m_reestimate(self, obs, alphabet, xi_list, epsilon_list, prob_list):

T = len(obs)
xi_list = deepcopy(xi_list)
epsilon_list = deepcopy(epsilon_list)

pi = {}
trans = {}
emis = {}

# Initial matrix

for i in range(0, len(states)): pi[states[i]] = xi_list[0][0][i]

# Transition matrix
    for i in range(0, len(states)):
    trans[states[i]] = {}

    for j in range(0, len(states)):
        above = 0
        below = 0

        for epsilon in epsilon_list:
            index = epsilon_list.index(epsilon)

            for t in range(1, len(obs[index])):

                above += epsilon[t][states[i]][j]

                for xi in xi_list:
                index = xi_list.index(xi)
                for t in range(0, len(obs[index])): 

                    below += xi[t][i]

            if below == 0: trans[states[i]][states[j]] = 0          
            else: trans[states[i]][states[j]] = above / below

    # Emission matrix
    for j in range(0, len(states)):
        emis[states[j]] = []

        for char in alphabet:
            above = 0
            below = 0

            for xi in xi_list:
                index = xi_list.index(xi)
                for t in range(0, len(xi)):

                    if obs[index][t] == char: above += xi[t][j]
                    below += xi[t][j]


            if below == 0: emis_t = 0 
            else: emis_t = float(above) / below
            emis[states[j]].append(emis_t)

    return {"init": init, "trans": trans, "emis": emis}

隐马尔可夫模型多重观测

0 个答案: