python语音模式检测内含数据集以及源码

# class Vad import numpy as np from scipy.signal import medfilt class VAD: def zc2(self, y, fn): """ short-term average zero-cross :param y: signal :param fn: frame number :return zcr: zero-cross """ if y.shape[1] != fn: y = y.T wlen = y.shape[0] # frame length zcr = np.zeros(fn) # initialization delta = 0.01 # small threshold for i in range(fn): yn = y[:, i] ym = np.zeros(len(yn)) for k in range(wlen): if yn[k] >= delta: ym[k] = yn[k] - delta elif yn[k] < -delta: ym[k] = yn[k] + delta else: ym[k] = 0 zcr[i] = np.sum(ym[0: -1] * ym[1: len(ym)] < 0) return zcr def findSegment(self, express): """ find voice start, end and length :param express: speech index :return soundSegment: """ express = express[0] if express[0] == 0 and express[1] == 0: # find where express = 1 voiceIndex = np.where(express) else: voiceIndex = express soundSegment = {} soundSegment.setdefault('begin', []).append(voiceIndex[0]) k = 1 for i in range(len(voiceIndex) - 1): if voiceIndex[i + 1] - voiceIndex[i] > 1: soundSegment.setdefault('end', []).append(voiceIndex[i]) soundSegment.setdefault('begin', []).append(voiceIndex[i + 1]) k = k + 1 soundSegment.setdefault('end', []).append(voiceIndex[-1]) for i in range(k): duration = soundSegment['end'][i] - soundSegment['begin'][i] + 1 soundSegment.setdefault('duration', []).append(duration) return soundSegment def multimidfilter(self, x, m): """ Multiple calls medfilt :param x: signal :param m: call times :return y: """ a = x for k in range(m): b = medfilt(a, 5) a = b y = a return y def vad_param1D(self, dst1, T1, T2): """ VAD with one parameter :param dst1: :param T1: threshold :param T2: :return voiceseg: :return vsl: :return SF: :return NF: """ dst1 = dst1.reshape(-1, 1) fn = dst1.shape[0] # frame number maxsilence = 8 # initialization minlen = 5 status = 0 count = [0] silence = [0] # Endpoint Detection xn = 0 x1 = [0] x2 = [0] for n in range(1, fn): if status == 0 or status == 1: # 0:silence, 1: maybe start if dst1[n] > T2: x1[xn] = np.max([n - count[xn], 1]) status = 2 silence[xn] = 0 count[xn] = count[xn] + 1 elif dst1[n] > T1: # maybe voice segment status = 1 count[xn] = count[xn] + 1 else: status = 0 # silence count[xn] = 0 x1[xn] = 0 x2[xn] = 0 elif status == 2: # voice if dst1[n] > T1: count[xn] = count[xn] + 1 else: silence[xn] = silence[xn] + 1 if silence[xn] < maxsilence: # silence exists, voice hasn't ended count[xn] = count[xn] + 1 elif count[xn] < minlen: # voice length too short, considered silent or noisy status = 0 silence[xn] = 0 count[xn] = 0 else: # voice end status = 3 x2[xn] = x1[xn] + count[xn] elif status == 3: # voice end, prepare for next status = 0 xn = xn + 1 count.append(0) silence.append(0) x1.append(0) x2.append(0) el = len(x1) if x1[el - 1] == 0: el = el - 1 # x1 real length if x2[el - 1] == 0: print('Error: Not find endding point! \n') x2[el - 1] = fn SF = np.zeros(fn) # define SF & NF, according to x1, x2 NF = np.ones(fn) for i in range(el): SF[(x1[i] - 1): x2[i]] = 1 NF[(x1[i] - 1): x2[i]] = 0 SpeechIndex = np.where(SF == 1) # voice segemnt vad = VAD() voiceseg = vad.findSegment(SpeechIndex) vsl = len(voiceseg['begin']) return voiceseg, vsl, SF, NF def vad_param1D_revr(self, dst1, T1, T2): """ VAD with one parameter reverse threshold :param dst1: :param T1: threshold :param T2: :return voiceseg: :return vsl: :return SF: :return NF: """ dst1 = dst1.reshape(-1, 1) fn = dst1.shape[0] # frame number maxsilence = 8 # initialization minlen = 5 status = 0 count = [0] silence = [0] # Endpoint Detection xn = 0 x1 = [0] x2 = [0] for n in range(1, fn): if status == 0 or status == 1: # 0:silence, 1: maybe start if dst1[n] < T2: x1[xn] = np.max([n - count[xn], 1]) status = 2 silence[xn] = 0 count[xn] = count[xn] + 1 elif dst1[n] < T1: # maybe voice segment status = 1 count[xn] = count[xn] + 1 else: status = 0 # silence count[xn] = 0 x1[xn] = 0 x2[xn] = 0 elif status == 2: # voice if dst1[n] < T1: count[xn] = count[xn] + 1 else: silence[xn] = silence[xn] + 1 if silence[xn] < maxsilence: # silence exists, voice hasn't ended count[xn] = count[xn] + 1 elif count[xn] < minlen: # voice length too short, considered silent or noisy status = 0 silence[xn] = 0 count[xn] = 0 else: # voice end status = 3 x2[xn] = x1[xn] + count[xn] elif status == 3: # voice end, prepare for next status = 0 xn = xn + 1 count.append(0) silence.append(0) x1.append(0) x2.append(0) el = len(x1) if x1[el - 1] == 0: el = el - 1 # x1 real length if x2[el - 1] == 0: print('Error: Not find endding point! \n') x2[el - 1] = fn SF = np.zeros(fn) # define SF & NF, according to x1, x2 NF = np.ones(fn) for i in range(el): SF[(x1[i] - 1): x2[i]] = 1 NF[(x1[i] - 1): x2[i]] = 0 SpeechIndex = np.where(SF == 1) # voice segemnt vad = VAD() voiceseg = vad.findSegment(SpeechIndex) vsl = len(voiceseg['begin']) return voiceseg, vsl, SF, NF def vad_param2D_revr(self, dst1, dst2, T1, T2, T3, T4=None): """ VAD with two parameters, dst1 & dst2, dst2: reverse comparison :param dst1: parameter1 :param dst2: parameter2 :param T1, T2, T3, T4: threshold :return voiceseg: :return vsl: :return SF: :return NF: """ fn = len(dst1) # frame number maxsilence = 8 # initialization minlen = 8 status = 0 count = [0] silence = [0] # Endpoint Detection xn = 0 x1 = [0] x2 = [0] for n in range(fn): if status == 0 or status == 1: # 0:silence, 1: maybe start if dst1[n] > T2 or (T4 != None and dst2[n] < T4): x1[xn] = np.max([n - count[xn], 1]) status = 2 silence[xn] = 0 count[xn] = count[xn] + 1 elif dst1[n] > T1 or dst2[n] < T3: # maybe voice segment status = 1 count[xn] = count[xn] + 1 else: status = 0 # silence count[xn] = 0 x1[xn] = 0 x2[xn] = 0 elif status == 2: # voice if dst1[n] > T1 or dst2[n] < T3: count[xn] = count[xn] + 1 else: silence[xn] = silence[xn] + 1 if silence[xn] < maxsilence: # silence exists, voice hasn't ended count[xn] = count[xn] + 1 elif count[xn] < minlen: # voice length too short, considered silent or noisy status = 0 silence[xn] = 0 count[xn] = 0 else: # voice end status = 3 x2[xn] = x1[xn] + count[xn] elif status == 3: # voice end, prepare for next status = 0 xn = xn + 1 count.append(0) silence.append(0) x1.append(0) x2.append(0) el = len(x1) if x1[el - 1] == 0: el = el - 1 # x1 real length if x2[el - 1] == 0: print('Error: Not find endding point! \n') x2[el - 1] = fn SF = np.zeros(fn) # define SF & NF, according to x1, x2 NF = np.ones(fn) for i in range(el): SF[(x1[i] - 1): x2[i]] = 1 NF[(x1[i] - 1): x2[i]] = 0 SpeechIndex = np.where(SF == 1) # voice segemnt vad = VAD() voiceseg = vad.findSegment(SpeechIndex) vsl = len(voiceseg['begin']) return voiceseg, vsl, SF, NF def vad(self, signal, noise, NoiseCounter=0, NoiseMargin=3.0, Hangover=8): """ Spectral Distance Voice Activity Detector :param signal: the current frames magnitude spectrum which is to labeld as noise or speech :param noise: noise magnitude spectrum template (estimation) :param NoiseCounter: the number of imediate previous noise frames :param NoiseMargin: