def filterByBasefrq(src, basefrq, width):
peaksPos = findpeaks(src, spacing=50, limit=max(src) * 0.05)
peaks = src[peaksPos] # 峰值大小
tar = np.copy(src)
num = min(int(len(src) / basefrq), 30)
for i in np.arange(num):
frq = i * basefrq
tar[frq - width:frq + width] = min(src[frq - width], src[frq + width])
return tar
def subpeakAmpLimiting(dataClip, space, limit):
peaks = findpeaks(dataClip, spacing=space, limit=max(dataClip) * limit)
if len(peaks) == 0:
return dataClip
dots = np.copy(dataClip[peaks])
dots[np.argmax(dots)] = 0
maxval = max(dots)
# 如果等于0 说明只有一个峰
if maxval > 0:
# dataClip = 3
np.where(dataClip < maxval, dataClip, maxval)
return dataClip
def subpeak_amplimiting(dataclip, space, limit):
"""
:param dataclip:输入
:param space:寻峰宽度
:param limit:寻峰阈值
:return:次高峰值限制幅度结果
"""
peaks = findpeaks(dataclip, spacing=space, limit=max(dataclip) * limit)
if len(peaks) == 0:
return dataclip
dots = np.copy(dataclip[peaks])
dots[np.argmax(dots)] = 0
maxval = max(dots)
# 如果等于0 说明只有一个峰
if maxval > 0:
dataclip = np.where(dataclip < maxval, dataclip, maxval)
return dataclip
def getPitch(dataClip,Fs,nfft):
pitch=0
dataClip[0:int(30*nfft/Fs)]=0
#print(frame*nfft/Fs) #当前时刻
fftData=fft(dataClip)[0:int(len(dataClip)/2)]
if showTestView:
plt.subplot(121)
plt.plot(np.arange(len(dataClip)), dataClip)
cepstrum=MaxMinNormalization(np.abs(fftData)*2/nfft,0,1)
cutoff=int(Fs/2/1200)
#cepstrum[0:cutoff]=0
cepstrum=MaxMinNormalization(cepstrum,0,1)
if showTestView:
plt.subplot(122)
plt.plot(np.arange(len(cepstrum)), cepstrum)
length=len(cepstrum)
peakind = signal.find_peaks_cwt(cepstrum, np.arange(1,20))
peakind2=findpeaks(cepstrum, spacing=15, limit=0.05)
if showTestView==1:
print(peakind)
print(peakind2)
# 把 corlor 设置为空，通过edgecolors来控制颜色
if showTestView==1:
plt.scatter(peakind2, cepstrum[peakind2], color='', marker='o', edgecolors='r', s=100)
peaks=[]
pos=cutoff
maxpos=np.argmax(cepstrum[pos:-1])+pos#最高峰位置
#检测半峰
simiPos=int(maxpos/2)
rad=int(simiPos/2)
simiMax=np.argmax(cepstrum[simiPos-rad:simiPos+rad])+simiPos-rad#半峰位置
bias=abs(simiPos-simiMax)/rad#偏离度10%
if bias<0.2:
#半峰与最高峰之间的最小值
simiMin=np.min(cepstrum[simiMax:maxpos])
#与半峰最小值差值
dist=cepstrum[simiMax]-simiMin
#差值比例
dist=dist/(cepstrum[maxpos]-simiMin)
if(dist>0.66):
#1 2
#return Fs/2/simiMax//
#四分检测
fourPos1=int(simiMax/2)#四分之一峰
p0=fourPos1
Xi=np.array([2,4])
Yi=np.array([simiMax,maxpos])
para=leastsq(error,p0,args=(Xi,Yi)) #把error函数中除了p以外的参数打包到args中
fourPos1=int(para[0])
fourPos2=int(para[0]*3)
rad=int(fourPos1/2)
fourMax1=np.argmax(cepstrum[fourPos1-rad:fourPos1+rad])+fourPos1-rad#四分峰位置1
bias=abs(fourPos1-fourMax1)/rad#偏离度10%
if bias<0.2:
#半峰与四分峰之间的最小值
fouMin1=np.min(cepstrum[fourMax1:simiMax])
#与半峰最小值差值
dist=cepstrum[fourMax1]-fouMin1
#差值比例
dist=dist/(cepstrum[simiMax]-fouMin1)
if(dist>0.66):
#寻找四分3峰
fourMax2=np.argmax(cepstrum[fourPos2-rad:fourPos2+rad])+fourPos2-rad#四分峰位置1
bias=abs(fourPos2-fourMax2)/rad#偏离度10%
if bias<0.2:
#添加四分峰
peaks.append([len(peaks)+1,fourMax1])
peaks.append([len(peaks)+1,simiMax])
peaks.append([len(peaks)+1,fourMax2])
if len(peaks)<1:
peaks.append([len(peaks)+1,simiMax])
#不存在半峰才检测三分峰
if len(peaks)<1:
thd1=0#第一个三分峰的位置
thd2=0#第二个三分峰的位置
#三分峰值检测
thrPos=int(maxpos*2/3)
rad=int((maxpos-simiPos)/2)
thrMax=np.argmax(cepstrum[thrPos-rad:thrPos+rad])+thrPos-rad#三分峰位置
bias=abs(thrPos-thrMax)/rad#偏离度10%
if bias<0.2:
#2/3峰与最大峰之间的最小数
thrMin=np.min(cepstrum[thrMax:maxpos])
#与最小值差值
dist=cepstrum[thrMax]-thrMin
#差值比例
dist=dist/(cepstrum[maxpos]-thrMin)
if(dist>0.66):
#return Fs/2/simiMax
#寻找1 如果寻找不到1 则取消2
thd1=thrMax
#三分峰2检测 （做回归）
p0=maxpos/3
Xi=np.array([2,3])
Yi=np.array([thd1,maxpos])
para=leastsq(error,p0,args=(Xi,Yi)) #把error函数中除了p以外的参数打包到args中
thrPos2=int(para[0]*1)
rad=int((thd1-thrPos2)/2)
#print(rad)
#print(thrPos2)
thrMax2=np.argmax(cepstrum[thrPos2-rad:thrPos2+rad])+thrPos2-rad#三分峰位置
bias=abs(thrPos2-thrMax2)/rad#偏离度10%
#print(bias)
if bias<0.2:
#与2/3峰之间的最小值
thrMin2=np.min(cepstrum[thrMax2:thrMax])
#与最小值差值
dist=cepstrum[thrMax]-thrMin2
#差值比例
dist=dist/(cepstrum[thrMax2]-thrMin2)
if(dist>0.66):
thd2=thrMax2
else:
thd1=0
thd2=0
else:
#1:
#return Fs/2/maxpos
thd1=0
thd2=0
if thd1>0:
peaks.append([len(peaks)+1,thd2])
peaks.append([len(peaks)+1,thd1])
#print(thd1)
#print(thd2)
peaks.append([len(peaks)+1,maxpos])
#每做一次线性回归 找下一个峰，直至找出右侧所有峰位置
getNextPeaks(peaks,cepstrum,0.1,0.25,0)#递归求右侧波峰
lenPeaks=len(peaks)
if showTestView:
if lenPeaks>0:
print(peaks)
rs=[]
if lenPeaks>1:
#线性拟合求基频，根据找到的所有二次fft峰求二次fft的基频，此基频求倒数然后乘以采样率就是原始信号的基频，对原始信号的高频部分分辨率不足
#，但原始信号高频部分在第一次fft中有好的频率分辨率，因此后期可以弥补
#harmonicIDs=np.arange(lenPeaks)+1#设置波峰ID
p0=peaks[lenPeaks-1][1]/lenPeaks#初始化参数
Xi=np.array(peaks)[:,0]
Yi=np.array(peaks)[:,1]
para=leastsq(error,p0,args=(Xi,Yi)) #把error函数中除了p以外的参数打包到args中
peaksfindWidth=int(nfft/2/para[0]*0.6)
peakindofFFT=findpeaks(dataClip, spacing=peaksfindWidth, limit=max(dataClip[int(30*nfft/Fs):-1])/10)
if showTestView==1:
plt.subplot(121)
plt.scatter(peakindofFFT, dataClip[peakindofFFT], color='', marker='o', edgecolors='r', s=100)
plt.show()
rs= [Fs/2/para[0],Fs/2/peaks[0][1]]
else:
if lenPeaks==1:
if showTestView==1:
plt.show()
rs= [0,0]
else:
plt.show()
rs= [0,0]
return rs
def getPitchDeScan(dataClip, Fs, nfft, showTestView):
pitch = 0
dataClip[0:int(30 * nfft / Fs)] = 0
dataClip = subpeakAmpLimiting(dataClip, int(30.0 / Fs * nfft), 0.1) # 次峰限幅
if showTestView:
plt.subplot(231)
plt.plot(np.arange(len(dataClip)), dataClip, label='amp-frq')
lowCutoff = int(40 * 441000.0 / Fs) # 最低截止频率对应的坐标
highCutoff = int(1400 * 441000.0 / Fs) # 最高截止频率对应的坐标
peakSearchPixes = int(3 * 441000 / Fs) # 寻峰间距
peakSearchAmp = 0.1 # 寻峰高度
# 线性内插重新采样
processingX = np.arange(0, min(int(
nfft / Fs * 4000), len(dataClip))) # 最大采集到4000Hz,不包括最大值，此处为尚未重采样的原始频谱
processingY = dataClip[processingX] # 重采样的fft
lenProcessingX = len(processingX) # 待处理频谱长度
finterp = interp1d(processingX, processingY, kind='linear') # 线性内插配置
x_pred = np.linspace(
0, processingX[lenProcessingX - 1] * 1.0,
int(processingX[lenProcessingX - 1] * 441000 / nfft) + 1)
maxProcessingX = x_pred[len(x_pred) - 1]
resampY = finterp(x_pred)
lenResampY = len(resampY)
# print(len(resampY))
if showTestView == 1:
plt.subplot(232)
plt.plot(np.arange(len(resampY)), resampY, label='rs_Amp-frq')
maxResampY = max(resampY[lowCutoff:-1]) / 2 # 待测频率内的最大值
# 测试梳状变换
# pixes=10
num = highCutoff # 栅栏变换后的长度
combTrans = np.zeros(num) # 存放栅栏变换的结果
indexComb = np.arange(lowCutoff, highCutoff, 0.1) # 栅栏变换索引
for k in np.arange(1, highCutoff, 1):
combTrans[k] = np.sum(
[resampY[i] for i in np.arange(0, lenResampY - 1, k)])
if showTestView == 1:
plt.subplot(233)
plt.plot(np.arange(len(combTrans)), combTrans, label='combTrans')
deSamp = [combTrans[m]
for m in np.arange(0, len(combTrans), 10)] # 10倍数降采样
deSamp[0] = 1000
if showTestView == 1:
plt.subplot(234)
plt.plot(np.arange(len(deSamp)), deSamp, label='DsCombTrans')
baseLine = getBaseLineFromScan(deSamp, num) # 通过扫描线算法求基准曲线
if showTestView == 1:
plt.subplot(235)
plt.plot(np.arange(len(baseLine)), baseLine, label='baseline')
trueTrans = combTrans - baseLine
trueTrans[0:lowCutoff] = 0
if showTestView == 1:
plt.subplot(236)
plt.plot(np.arange(len(trueTrans)), trueTrans, label='trueTrans')
if (sum(dataClip) < 1):
return [0 / 10.0, resampY, trueTrans]
pitch = max(trueTrans) / sum(dataClip)
# 频率采样变换后寻峰
combTransPeaks = findpeaks(trueTrans,
spacing=peakSearchPixes,
limit=max(trueTrans) * peakSearchAmp)
peaks = trueTrans[combTransPeaks] # 峰值大小
if (len(peaks) == 0):
return [0 / 10.0, resampY, trueTrans]
maxindex = np.argmax(peaks)
maxfrq = combTransPeaks[maxindex] # 最高峰值位置
# 寻找1hz以内的最大的峰
combThr = 6 * 441000 / Fs # 寻峰宽度
decThr = 0.65 # 递减阈值
preBaseFrq = maxfrq
for n in range(2, 10, 1):
newfrq = getNearPeaks(trueTrans, combTransPeaks, peaks, n * maxfrq,
combThr, preBaseFrq, decThr, showTestView)
if (newfrq > 0):
preBaseFrq = newfrq
else:
continue
pitch = preBaseFrq
if showTestView == 1:
plt.scatter(combTransPeaks,
trueTrans[combTransPeaks],
color='',
marker='o',
edgecolors='r',
s=100)
plt.show()
return [pitch / 10.0, resampY, trueTrans]
def getpitch(self, dataclip, fs, nfft, showtestview):
"""
音高估计
:param dataclip:输入fft
:param fs:采样率
:param nfft:窗口大小
:param showtestview:是否显示figure
:return:频率,插值采样后的输入,梳状统计向量
"""
dataclip[0:int(30 * nfft / fs)] = 0
dataclip = BaseFrqDetector.subpeak_amplimiting(dataclip,
int(30.0 / fs * nfft),
0.1) # 次峰限幅
lowcutoff = int(32.5 * 441000.0 / fs) # 最低截止频率对应的坐标
highcutoff = int(1400 * 441000.0 / fs) # 最高截止频率对应的坐标
peaksearchpixes = int(3 * 441000 / fs) # 寻峰间距
peaksearchamp = 0.1 # 寻峰高度
# 线性内插重新采样
processingx = np.arange(
0, min(int(nfft / fs * 4000),
len(dataclip))) # 最大采集到4000Hz,不包括最大值，此处为尚未重采样的原始频谱
processingy = dataclip[processingx] # 重采样的fft
lenprocessingx = len(processingx) # 待处理频谱长度
finterp = interp1d(processingx, processingy, kind='linear') # 线性内插配置
x_pred = np.linspace(
0, processingx[lenprocessingx - 1] * 1.0,
int(processingx[lenprocessingx - 1] * 441000 / nfft) + 1)
resampy = finterp(x_pred)
lenresampy = len(resampy)
# pixes=10
num = highcutoff # 栅栏变换后的长度
combtrans = np.zeros(num) # 存放栅栏变换的结果
# 梳状变换, 2019-03-10 16:54:45
for k in np.arange(lowcutoff, highcutoff, 1):
combtrans[k] = sum(resampy[::k])
combtrans[0:lowcutoff] = max(combtrans)
# 如果有去扫描参数,则去扫描
if self.isdescan is True:
# 用于降低采样
desamp = [combtrans[m]
for m in np.arange(0, len(combtrans), 10)] # 10倍数降采样
desamp[0] = np.max(desamp)
baseline = BaseFrqDetector.getbaselinefromscan(desamp,
num) # 通过扫描线算法求基准曲线
truetrans = combtrans - baseline
else:
truetrans = combtrans
truetrans[0:lowcutoff] = 0
if sum(dataclip) < 1:
return [0 / 10.0, resampy, truetrans]
# 频率采样变换后寻峰
combtranspeaks = findpeaks(truetrans,
spacing=peaksearchpixes,
limit=max(truetrans) * peaksearchamp)
peaks = truetrans[combtranspeaks] # 峰值大小
if len(peaks) == 0:
return [0 / 10.0, resampy, truetrans]
maxindex = np.argmax(peaks)
maxfrq = combtranspeaks[maxindex] # 最高峰值位置
# 寻找1hz以内的最大的峰
combthr = 6 * 441000 / fs # 寻峰宽度
decthr = 0.65 # 递减阈值
prebasefrq = maxfrq
for n in range(2, 10, 1):
newfrq = BaseFrqDetector.getnearpeaks(truetrans, combtranspeaks,
peaks, n * maxfrq, combthr,
prebasefrq, decthr,
showtestview)
if newfrq > 0:
prebasefrq = newfrq
else:
continue
pitch = prebasefrq
return [pitch / 10.0, resampy, truetrans]
