python学习之实现语音的简单训练及识别

语音识别是一门交叉学科。近二十年来，语音识别技术取得显著进步，开始从实验室走向市场。人们预计，未来10年内，语音识别技术将进入工业、家电、通信、汽车电子、医疗、家庭服务、消费电子产品等各个领域。 语音识别听写机在一些领域的应用被美国新闻界评为1997年计算机发展十件大事之一。很多专家都认为语音识别技术是2000年至2010年间信息技术领域十大重要的科技发展技术之一。 语音识别技术所涉及的领域包括：信号处理、模式识别、概率论和信息论、发声机理和听觉机理、人工智能等等。

此次用的是python3.6的版本实现的，安装一些相关模块sklearn:

pip install sklearn

会自动下载相关的包，好像一般来说，是下载最新的（不敢确认）。

中间也可能会存在一定的错误，导致失败，但可以重新，继续下载，反正我是这样的。

知道成功安装。

gParam的代码：

#!/usr/bin/env python3
# -*- coding:utf-8 -*-
u'''
Created on 2019年4月27日@author: wuluo
'''
__author__ = 'wuluo'
__version__ = '1.0.0'
__company__ = u'重庆交大'
__updated__ = '2019-04-27'TRAIN_DATA_PATH = 'G:/2018and2019two/duomeitijishu/data/train/'
TEST_DATA_PATH = 'G:/2018and2019two/duomeitijishu/data/test/'
NSTATE = 4
NPDF = 3
MAX_ITER_CNT = 100
NUM = 10if __name__ == "__main__":pass

对于读入的路径，我用的是绝对路径，也就是写死。
test的代码：

#!/usr/bin/env python3
# -*- coding:utf-8 -*-
u'''
Created on 2019年4月27日@author: wuluo
'''
__author__ = 'wuluo'
__version__ = '1.0.0'
__company__ = u'重庆交大'
__updated__ = '2019-04-27'import numpy as np
from numpy import *
from duomeiti import gParam
from duomeiti import my_hmm
from my_hmm import gmm_hmm  #可会出现报错，但实际运行起来，没有问题
my_gmm_hmm = gmm_hmm()my_gmm_hmm.loadWav(gParam.TRAIN_DATA_PATH)
my_gmm_hmm.hmm_start_train()
my_gmm_hmm.recog(gParam.TEST_DATA_PATH)if __name__ == "__main__":pass

在编译与修改代码的边缘，疯狂试探；

主要还是python版本的原因，版本的不同，有些函数的用法也不同。具体详细请见这个网址：https://www.runoob.com/python3/python3-att-list-extend.html

my_hmm的代码：

#!/usr/bin/env python3
# -*- coding:utf-8 -*-
u'''
Created on 2019年4月27日@author: wuluo
'''
__author__ = 'wuluo'
__version__ = '1.0.0'
__company__ = u'重庆交大'
__updated__ = '2019-04-27'import numpy as np
from numpy import *
from sklearn.cluster import KMeans
from scipy import sparse
import scipy.io as sio
from scipy import signal
import wave
import math
from duomeiti import gParam
import copydef pdf(m, v, x):est_v = np.prod(v, axis=0)test_x = np.dot((x - m) / v, x - m)p = (2 * math.pi * np.prod(v, axis=0))**-0.5 * \np.exp(-0.5 * np.dot((x - m) / v, x - m))return pclass sampleInfo:def __init__(self):self.smpl_wav = []self.smpl_data = []self.seg = []def set_smpl_wav(self, wav):self.smpl_wav.append(wav)def set_smpl_data(self, data):self.smpl_data.append(data)def set_segment(self, seg_list):self.seg = seg_listclass mixInfo:def __init__(self):self.Cmean = []self.Cvar = []self.Cweight = []self.CM = []class hmmInfo:def __init__(self):self.init = []  # 初始矩阵self.trans = []  # 转移概率矩阵self.mix = []  # 高斯混合模型参数self.N = 0  # 状态数class gmm_hmm:def __init__(self):self.hmm = []  # 单个hmm序列,self.gmm_hmm_model = []  # 把所有的训练好的gmm-hmm写入到这个队列self.samples = []  # 0-9 所有的音频数据self.smplInfo = []  # 这里面主要是单个数字的音频数据和对应mfcc数据# 每一个HMM对应len(stateInfo)个状态,每个状态指定高斯个数(3)self.stateInfo = [gParam.NPDF, gParam.NPDF, gParam.NPDF, gParam.NPDF]def loadWav(self, pathTop):for i in range(gParam.NUM):tmp_data = []for j in range(gParam.NUM):wavPath = pathTop + str(i) + str(j) + '.wav'f = wave.open(wavPath, 'rb')params = f.getparams()nchannels, sampwidth, framerate, nframes = params[:4]str_data = f.readframes(nframes)# print shape(str_data)f.close()wave_data = np.fromstring(str_data, dtype=short) / 32767.0#wave_data.shape = -1,2#wave_data = wave_data.T#wave_data = wave_data.reshape(1,wave_data.shape[0]*wave_data.shape[1])# print shape(wave_data),type(wave_data)tmp_data.append(wave_data)self.samples.append(tmp_data)# 循环读数据,然后进行训练def hmm_start_train(self):Nsmpls = len(self.samples)for i in range(Nsmpls):tmpSmplInfo0 = []n = len(self.samples[i])for j in range(n):tmpSmplInfo1 = sampleInfo()tmpSmplInfo1.set_smpl_wav(self.samples[i][j])tmpSmplInfo0.append(tmpSmplInfo1)# self.smplInfo.append(tmpSmplInfo0)print('现在训练第' + str(i) + '个HMM模型')hmm0 = self.trainhmm(tmpSmplInfo0, self.stateInfo)print('第' + str(i) + '个模型已经训练完毕')# self.gmm_hmm_model.append(hmm0)# 训练hmmdef trainhmm(self, sample, state):K = len(sample)print('首先进行语音参数计算-MFCC')for k in range(K):tmp = self.mfcc(sample[k].smpl_wav)sample[k].set_smpl_data(tmp)  # 设置MFCCdatahmm = self.inithmm(sample, state)pout = zeros((gParam.MAX_ITER_CNT, 1))for my_iter in range(gParam.MAX_ITER_CNT):print('第' + str(my_iter) + '训练')hmm = self.baum(hmm, sample)for k in range(K):pout[my_iter, 0] = pout[my_iter, 0] + \self.viterbi(hmm, sample[k].smpl_data[0])if my_iter > 0:if(abs((pout[my_iter, 0] - pout[my_iter - 1, 0]) / pout[my_iter, 0]) < 5e-6):print('收敛')self.gmm_hmm_model.append(hmm)return hmmself.gmm_hmm_model.append(hmm)# 获取MFCC参数def mfcc(self, k):M = 24  # 滤波器的个数N = 256  # 一帧语音的采样点数arr_mel_bank = self.melbank(M, N, 8000, 0, 0.5, 'm')arr_mel_bank = arr_mel_bank / np.amax(arr_mel_bank)#计算DCT系数, 12*24rDCT = 12cDCT = 24dctcoef = []for i in range(1, rDCT + 1):tmp = [np.cos((2 * j + 1) * i * math.pi * 1.0 / (2.0 * cDCT))for j in range(cDCT)]dctcoef.append(tmp)# 归一化倒谱提升窗口w = [1 + 6 * np.sin(math.pi * i * 1.0 / rDCT)for i in range(1, rDCT + 1)]w = w / np.amax(w)# 预加重AggrK = double(k)AggrK = signal.lfilter([1, -0.9375], 1, AggrK)  # ndarray#AggrK = AggrK.tolist()# 分帧FrameK = self.enframe(AggrK[0], N, 80)n0, m0 = FrameK.shapefor i in range(n0):#temp = multiply(FrameK[i,:],np.hamming(N))# print shape(temp)FrameK[i, :] = multiply(FrameK[i, :], np.hamming(N))FrameK = FrameK.T# 计算功率谱S = (abs(np.fft.fft(FrameK, axis=0)))**2# 将功率谱通过滤波器组P = np.dot(arr_mel_bank, S[0:129, :])# 取对数后做余弦变换D = np.dot(dctcoef, log(P))n0, m0 = D.shapem = []for i in range(m0):m.append(np.multiply(D[:, i], w))n0, m0 = shape(m)dtm = zeros((n0, m0))for i in range(2, n0 - 2):dtm[i, :] = -2 * m[i - 2][:] - m[i - 1][:] + \m[i + 1][:] + 2 * m[i + 2][:]dtm = dtm / 3.0# cc = [m,dtm]cc = np.column_stack((m, dtm))# cc.extend(list(dtm))cc = cc[2:n0 - 2][:]# print shape(cc)return cc# melbankdef melbank(self, p, n, fs, f1, fh, w):f0 = 700.0 / (1.0 * fs)fn2 = floor(n / 2.0)lr = math.log((float)(f0 + fh) / (float)(f0 + f1)) / (float)(p + 1)tmpList = [0, 1, p, p + 1]bbl = []for i in range(len(tmpList)):bbl.append(n * ((f0 + f1) * math.exp(tmpList[i] * lr) - f0))#b1 = n*((f0+f1) * math.exp([x*lr for x in tmpList]) - f0)# print bblb2 = ceil(bbl[1])b3 = floor(bbl[2])if(w == 'y'):pf = np.log((f0 + range(b2, b3) / n) / (f0 + f1)) / lr  # notefp = floor(pf)r = [ones((1, b2)), fp, fp + 1, p * ones((1, fn2 - b3))]c = [range(0, b3), range(b2, fn2)]v = 2 * [0.5, ones((1, b2 - 1)), 1 - pf + fp,pf - fp, ones((1, fn2 - b3 - 1)), 0.5]mn = 1mx = fn2 + 1else:b1 = floor(bbl[0]) + 1b4 = min([fn2, ceil(bbl[3])]) - 1pf = []for i in range(int(b1), int(b4 + 1), 1):pf.append(math.log((f0 + (1.0 * i) / n) / (f0 + f1)) / lr)fp = floor(pf)pm = pf - fpk2 = b2 - b1 + 1k3 = b3 - b1 + 1k4 = b4 - b1 + 1r = fp[int(k2 - 1):int(k4)]r1 = 1 + fp[0:int(k3)]r = r.tolist()r1 = r1.tolist()r.extend(r1)#r = [fp[int(k2-1):int(k4)],1+fp[0:int(k3)]]c = list(range(int(k2), int(k4 + 1)))c2 = list(range(1, int(k3 + 1)))# c = c.tolist()# c2 = c2.tolist()c.extend(c2)#c = [range(int(k2),int(k4+1)),range(0,int(k3))]v = 1 - pm[int(k2 - 1):int(k4)]v = v.tolist()v1 = pm[0:int(k3)]v1 = v1.tolist()v.extend(v1)  # [1-pm[int(k2-1):int(k4)],pm[0:int(k3)]]v = [2 * x for x in v]mn = b1 + 1mx = b4 + 1if(w == 'n'):v = 1 - math.cos(v * math.pi / 2)elif (w == 'm'):tmpV = []# for i in range(v):#     tmpV.append(1-0.92/1.08*math.cos(v[i]*math))v = [1 - 0.92 / 1.08 * math.cos(x * math.pi / 2) for x in v]# print type(c),type(mn)col_list = [x + int(mn) - 2 for x in c]r = [x - 1 for x in r]x = sparse.coo_matrix((v, (r, col_list)), shape=(p, 1 + int(fn2)))matX = x.toarray()#np.savetxt('./data.csv',matX, delimiter=' ')return matX  # x.toarray()# 分帧函数def enframe(self, x, win, inc):nx = len(x)try:nwin = len(win)except Exception as err:# print errnwin = 1if (nwin == 1):wlen = winelse:wlen = nwin# print inc,wlen,nx# here has a bug that nf maybe less than 0nf = fix(1.0 * (nx - wlen + inc) / inc)f = zeros((int(nf), wlen))indf = [inc * j for j in range(int(nf))]indf = (mat(indf)).Tinds = mat(range(wlen))indf_tile = tile(indf, wlen)inds_tile = tile(inds, (int(nf), 1))mix_tile = indf_tile + inds_tilefor i in range(int(nf)):for j in range(wlen):f[i, j] = x[mix_tile[i, j]]# print x[mix_tile[i,j]]if nwin > 1:  # TODOdw = win.tolist()#w_tile = tile(w,(int))return f# init hmmdef inithmm(self, sample, M):K = len(sample)N0 = len(M)self.N = N0# 初始概率矩阵hmm = hmmInfo()hmm.init = zeros((N0, 1))hmm.init[0] = 1hmm.trans = zeros((N0, N0))hmm.N = N0# 初始化转移概率矩阵for i in range(self.N - 1):hmm.trans[i, i] = 0.5hmm.trans[i, i + 1] = 0.5hmm.trans[self.N - 1, self.N - 1] = 1# 概率密度函数的初始聚类# 分段for k in range(K):T = len(sample[k].smpl_data[0])#seg0 = []seg0 = np.floor(arange(0, T, 1.0 * T / N0))#seg0 = int(seg0.tolist())seg0 = np.concatenate((seg0, [T]))# seg0.append(T)sample[k].seg = seg0# 对属于每个状态的向量进行K均值聚类,得到连续混合正态分布mix = []for i in range(N0):vector = []for k in range(K):seg1 = int(sample[k].seg[i])seg2 = int(sample[k].seg[i + 1])tmp = []tmp = sample[k].smpl_data[0][seg1:seg2][:]if k == 0:vector = np.array(tmp)else:vector = np.concatenate((vector, np.array(tmp)))# vector.append(tmp)# tmp_mix = mixInfo()# print id(tmp_mix)tmp_mix = self.get_mix(vector, M[i], mix)# mix.append(tmp_mix)hmm.mix = mixreturn hmm# get mix datadef get_mix(self, vector, K, mix0):kmeans = KMeans(n_clusters=K, random_state=0).fit(np.array(vector))# 计算每个聚类的标准差,对角阵,只保存对角线上的元素mix = mixInfo()var0 = []mean0 = []#ind = []for j in range(K):#ind = [i for i in kmeans.labels_ if i==j]ind = []ind1 = 0for i in kmeans.labels_:if i == j:ind.append(ind1)ind1 = ind1 + 1tmp = [vector[i][:] for i in ind]var0.append(np.std(tmp, axis=0))mean0.append(np.mean(tmp, axis=0))weight0 = zeros((K, 1))for j in range(K):tmp = 0ind1 = 0for i in kmeans.labels_:if i == j:tmp = tmp + ind1ind1 = ind1 + 1weight0[j] = tmpweight0 = weight0 / weight0.sum()mix.Cvar = multiply(var0, var0)mix.Cmean = mean0mix.CM = Kmix.Cweight = weight0mix0.append(mix)return mix0# baum-welch 算法实现函数体def baum(self, hmm, sample):mix = copy.deepcopy(hmm.mix)  # 高斯混合N = len(mix)  # HMM状态数K = len(sample)  # 语音样本数SIZE = shape(sample[0].smpl_data[0])[1]  # 参数阶数,MFCC维数print('计算样本参数.....')c = []alpha = []beta = []ksai = []gama = []for k in range(K):c0, alpha0, beta0, ksai0, gama0 = self.getparam(hmm, sample[k].smpl_data[0])c.append(c0)alpha.append(alpha0)beta.append(beta0)ksai.append(ksai0)gama.append(gama0)# 重新估算概率转移矩阵print('----- 重新估算概率转移矩阵 -----')for i in range(N - 1):denom = 0for k in range(K):ksai0 = ksai[k]tmp = ksai0[:, i, :]  # ksai0[:][i][:]denom = denom + sum(tmp)for j in range(i, i + 2):norm = 0for k in range(K):ksai0 = ksai[k]tmp = ksai0[:, i, j]  # [:][i][j]norm = norm + sum(tmp)hmm.trans[i, j] = norm / denom# 重新估算发射概率矩阵,即GMM的参数print('----- 重新估算输出概率矩阵,即GMM的参数 -----')for i in range(N):for j in range(mix[i].CM):nommean = zeros((1, SIZE))nomvar = zeros((1, SIZE))denom = 0for k in range(K):gama0 = gama[k]T = shape(sample[k].smpl_data[0])[0]for t in range(T):x = sample[k].smpl_data[0][t][:]nommean = nommean + gama0[t, i, j] * xnomvar = nomvar + gama0[t, i, j] * \(x - mix[i].Cmean[j][:])**2denom = denom + gama0[t, i, j]hmm.mix[i].Cmean[j][:] = nommean / denomhmm.mix[i].Cvar[j][:] = nomvar / denomnom = 0denom = 0# 计算pdf权值for k in range(K):gama0 = gama[k]tmp = gama0[:, i, j]nom = nom + sum(tmp)tmp = gama0[:, i, :]denom = denom + sum(tmp)hmm.mix[i].Cweight[j] = nom / denomreturn hmm# 前向-后向算法def getparam(self, hmm, O):T = shape(O)[0]init = hmm.init  # 初始概率trans = copy.deepcopy(hmm.trans)  # 转移概率mix = copy.deepcopy(hmm.mix)  # 高斯混合N = hmm.N  # 状态数# 给定观测序列,计算前向概率alphax = O[0][:]alpha = zeros((T, N))#----- 计算前向概率alpha -----#for i in range(N):  # t=0tmp = hmm.init[i] * self.mixture(mix[i], x)alpha[0, i] = tmp  # hmm.init[i]*self.mixture(mix[i],x)# 标定t=0时刻的前向概率c = zeros((T, 1))c[0] = 1.0 / sum(alpha[0][:])alpha[0][:] = c[0] * alpha[0][:]for t in range(1, T, 1):  # t = 1~Tfor i in range(N):temp = 0.0for j in range(N):temp = temp + alpha[t - 1, j] * trans[j, i]alpha[t, i] = temp * self.mixture(mix[i], O[t][:])c[t] = 1.0 / sum(alpha[t][:])alpha[t][:] = c[t] * alpha[t][:]#----- 计算后向概率 -----#beta = zeros((T, N))for i in range(N):  # T时刻beta[T - 1, i] = c[T - 1]for t in range(T - 2, -1, -1):x = O[t + 1][:]for i in range(N):for j in range(N):beta[t, i] = beta[t, i] + beta[t + 1, j] * \self.mixture(mix[j], x) * trans[i, j]beta[t][:] = c[t] * beta[t][:]# 过渡概率ksaiksai = zeros((T - 1, N, N))for t in range(0, T - 1):denom = sum(np.multiply(alpha[t][:], beta[t][:]))for i in range(N - 1):for j in range(i, i + 2, 1):norm = alpha[t, i] * trans[i, j] * \self.mixture(mix[j], O[t + 1][:]) * beta[t + 1, j]ksai[t, i, j] = c[t] * norm / denom# 混合输出概率 gamagama = zeros((T, N, max(self.stateInfo)))for t in range(T):pab = zeros((N, 1))for i in range(N):pab[i] = alpha[t, i] * beta[t, i]x = O[t][:]for i in range(N):prob = zeros((mix[i].CM, 1))for j in range(mix[i].CM):m = mix[i].Cmean[j][:]v = mix[i].Cvar[j][:]prob[j] = mix[i].Cweight[j] * pdf(m, v, x)if mix[i].Cweight[j] == 0.0:print(pdf(m, v, x))tmp = pab[i] / pab.sum()tmp = tmp[0]temp_sum = prob.sum()for j in range(mix[i].CM):gama[t, i, j] = tmp * prob[j] / temp_sumreturn c, alpha, beta, ksai, gamadef mixture(self, mix, x):prob = 0.0for i in range(mix.CM):m = mix.Cmean[i][:]v = mix.Cvar[i][:]w = mix.Cweight[i]tmp = pdf(m, v, x)# print tmpprob = prob + w * tmp  # * pdf(m,v,x)if prob == 0.0:prob = 2e-100return prob# 维特比算法def viterbi(self, hmm, O):init = copy.deepcopy(hmm.init)trans = copy.deepcopy(hmm.trans)  # hmm.transmix = hmm.mixN = hmm.NT = shape(O)[0]# 计算Log(init)n_init = len(init)for i in range(n_init):if init[i] <= 0:init[i] = -infelse:init[i] = log(init[i])# 计算log(trans)m, n = shape(trans)for i in range(m):for j in range(n):if trans[i, j] <= 0:trans[i, j] = -infelse:trans[i, j] = log(trans[i, j])# 初始化delta = zeros((T, N))fai = zeros((T, N))q = zeros((T, 1))# t=0x = O[0][:]for i in range(N):delta[0, i] = init[i] + log(self.mixture(mix[i], x))# t=2:Tfor t in range(1, T):for j in range(N):tmp = delta[t - 1][:] + trans[:][j].Ttmp = tmp.tolist()delta[t, j] = max(tmp)fai[t, j] = tmp.index(max(tmp))x = O[t][:]delta[t, j] = delta[t, j] + log(self.mixture(mix[j], x))tmp = delta[T - 1][:]tmp = tmp.tolist()prob = max(tmp)q[T - 1] = tmp.index(max(tmp))for t in range(T - 2, -1, -1):q[t] = fai[t + 1, int(q[t + 1, 0])]return prob# 用于测试的程序 def vad(self, k, fs):k = double(k)k = multiply(k, 1.0 / max(abs(k)))# 计算短时过零率FrameLen = 240FrameInc = 80FrameTemp1 = self.enframe(k[0:-2], FrameLen, FrameInc)FrameTemp2 = self.enframe(k[1:], FrameLen, FrameInc)signs = np.sign(multiply(FrameTemp1, FrameTemp2))signs = list(map(lambda x: [[i, 0][i > 0] for i in x], signs))signs = list(map(lambda x: [[i, 1][i < 0] for i in x], signs))diffs = np.sign(abs(FrameTemp1 - FrameTemp2) - 0.01)diffs = list(map(lambda x: [[i, 0][i < 0] for i in x], diffs))zcr = sum(multiply(signs, diffs), 1)# 计算短时能量amp = sum(abs(self.enframe(signal.lfilter([1, -0.9375], 1, k), FrameLen, FrameInc)), 1)# print '短时能量%f' %amp# 设置门限print('设置门限')ZcrLow = max([round(mean(zcr) * 0.1), 3])  # 过零率低门限ZcrHigh = max([round(max(zcr) * 0.1), 5])  # 过零率高门限AmpLow = min([min(amp) * 10, mean(amp) * 0.2, max(amp) * 0.1])  # 能量低门限AmpHigh = max([min(amp) * 10, mean(amp) *0.2, max(amp) * 0.1])  # 能量高门限# 端点检测MaxSilence = 8  # 最长语音间隙时间MinAudio = 16  # 最短语音时间Status = 0  # 状态0:静音段,1:过渡段,2:语音段,3:结束段HoldTime = 0  # 语音持续时间SilenceTime = 0  # 语音间隙时间print('开始端点检测')StartPoint = 0for n in range(len(zcr)):if Status == 0 or Status == 1:if amp[n] > AmpHigh or zcr[n] > ZcrHigh:StartPoint = n - HoldTimeStatus = 2HoldTime = HoldTime + 1SilenceTime = 0elif amp[n] > AmpLow or zcr[n] > ZcrLow:Status = 1HoldTime = HoldTime + 1else:Status = 0HoldTime = 0elif Status == 2:if amp[n] > AmpLow or zcr[n] > ZcrLow:HoldTime = HoldTime + 1else:SilenceTime = SilenceTime + 1if SilenceTime < MaxSilence:HoldTime = HoldTime + 1elif (HoldTime - SilenceTime) < MinAudio:Status = 0HoldTime = 0SilenceTime = 0else:Status = 3elif Status == 3:breakif Status == 3:breakHoldTime = HoldTime - SilenceTimeEndPoint = StartPoint + HoldTimereturn StartPoint, EndPointdef recog(self, pathTop):N = gParam.NUMfor i in range(N):wavPath = pathTop + str(i) + '.wav'f = wave.open(wavPath, 'rb')params = f.getparams()nchannels, sampwidth, framerate, nframes = params[:4]str_data = f.readframes(nframes)# print shape(str_data)f.close()wave_data = np.fromstring(str_data, dtype=short) / 32767.0x1, x2 = self.vad(wave_data, framerate)O = self.mfcc([wave_data])O = O[x1 - 3:x2 - 3][:]print('第' + str(i) + '个词的观察矢量是:' + str(i))pout = []for j in range(N):pout.append(self.viterbi(self.gmm_hmm_model[j], O))n = pout.index(max(pout))print('第' + str(i) + '个词,识别是:' + str(n))if __name__ == "__main__":pass

这里是运行结果：

本文来自互联网用户投稿，文章观点仅代表作者本人，不代表本站立场，不承担相关法律责任。如若转载，请注明出处。 如若内容造成侵权/违法违规/事实不符，请点击【内容举报】进行投诉反馈！