python语音处理：语音增强算法内含数据集以及源码

# # Class Enhancement from scipy.signal import lfilter from spectrum import pmtm from Universal import * from VAD import * class Enhancement: def simplesubspec(self, signal, wlen, inc, NIS, a, b): """ simple spectrum subtraction :param signal: noisy speech :param wlen: window length :param inc: frame shift :param NIS: leading noise segment length :param a: over subtraction factor :param b: gain factor :return output: denoise speech """ wnd = np.hamming(wlen) # window function N = len(signal) # signal length speech = Speech() y = speech.enframe(signal, list(wnd), inc).T # enframe fn = y.shape[1] # frame number y_fft = np.fft.fft(y, axis=0) # FFT y_a = np.abs(y_fft) # amplitude y_phase = np.angle(y_fft) # phase y_a2 = y_a ** 2 # energy Nt = np.mean(y_a2[:, 0: NIS], 1) # average energy in noise segment nl2 = int(wlen / 2) + 1 # positvie frequency temp = np.zeros(nl2) # energy U = np.zeros(nl2) # one frame amplitude X = np.zeros((nl2, fn)) # amplitude for i in range(fn): # spectrum subtraction for k in range(nl2): if (y_a2[k, i] > a * Nt[k]): temp[k] = y_a2[k, i] - a * Nt[k] else: temp[k] = b * y_a2[k, i] U[k] = np.sqrt(temp[k]) X[:, i] = U output = speech.OverlapAdd2(X, y_phase[0:nl2, :], wlen, inc) # synthesis Nout = len(output) # spectrum subtraction length = original length? if Nout > N: output = output[0: N] else: output = np.concatenate([output, np.zeros(N - Nout)]) output = output / np.max(np.abs(output)) # normalization return output def segment(self, siganl, W=256, SP=0.4, Window=np.hamming(256)): """ chops a signal to overlapping windowed segments :param siganl: one dimentional signal :param W: sample number per window (default = 25) :param SP: shift percent (default = 0.4) :param Window: window function (default: hamming) :return Seg: segment matrix """ if (W != 256): Window = np.hamming(W) Window = Window.reshape(-1, 1) # make it a column vector L = len(siganl) SP = int(W * SP) N = int((L - W) / SP + 1) # number of segments Index = np.tile(np.arange(0, W), (N, 1)) + np.tile(SP * np.arange(0, N).reshape(-1, 1), (1, W)) Index = Index.T hw = np.tile(Window, (1, N)) Seg = siganl[Index] * hw return Seg def SSBoll79(self, signal, fs, IS=None): """ Spectral Subtraction based on Boll 79. Amplitude spectral subtraction Includes Magnitude Averaging and Residual noise Reduction :param signal: noisy signal :param fs: sampling frequency :param IS: initial silence (noise only) length in seconds (default value is .25 sec) :return output: denoise signal """ if not IS: IS = 0.25 # seconds elif isinstance(IS, float): W = int(0.025 * fs) # window length 25ms nfft = W # overlap-add method works good with this shift value SP = 0.4 # frame shift 40% (10ms) wnd = np.hamming(W) # IGNORE THIS SECTION FOR COMPATIBILITY WITH ANOTHER PROGRAM FROM HERE..... if isinstance(IS, dict): W = IS['windowsize'] SP = IS['shiftsize'] / W nfft = IS['nfft'] wnd = IS['window'] if hasattr(IS, 'IS'): IS = IS['IS'] else: IS = 0.25 # .......IGNORE THIS SECTION FOR COMPATIBILITY WITH ANOTHER PROGRAM T0 HERE NIS = int((IS * fs - W) / (SP * W) + 1) # number of initial silence segments Gamma = 1 # 1: magnitude, 2: power spectrum y = self.segment(signal, W, SP, wnd) Y = np.fft.fft(y, axis=0) FreqResol, NumberofFrames = Y.shape YPhase = np.angle(Y[0: int(NumberofFrames / 2) + 1, :]) # noisy speech phase Y = np.abs(Y[0: int(NumberofFrames / 2) + 1, :]) ** Gamma # Spectrogram N = np.mean(Y[:, 0:NIS].T, axis=0).T # initial noise power spectrum mean NRM = np.zeros(N.shape) # Noise Residual Maximum (Initialization) NoiseCounter = 0 NoiseLength = 9 # smoothing factor for noise updating Beta = 0.03 YS = Y # Y magnitude average for i in np.arange(1, NumberofFrames - 1): YS[:, i] = (YS[:, i - 1] + YS[:, i] + YS[:, i + 1]) / 3 X = np.zeros(Y.shape) D = np.zeros(FreqResol) for i in range(NumberofFrames): # Magnitude Spectrum Distance VAD NoiseFlag, SpeechFlag, NoiseCounter, Dist = VAD().vad(Y[:, i] ** (1 / Gamma), N ** (1 / Gamma), NoiseCounter) if SpeechFlag == 0: N = (NoiseLength * N + Y[:, i]) / (NoiseLength + 1) # update and smooth noise NRM = np.maximum(NRM, YS[:, i] - N) # update maximum noise residue X[:, i] = Beta * Y[:, i] else: D = YS[:, i] - N # spectral subtraction if i > 0 and i < NumberofFrames - 1: # residual noise reduction for j in range(len(D)): if D[j] < NRM[j]: D[j] = np.min(np.array([D[j], YS[j, i - 1] - N[j], YS[j, i + 1] - N[j]])) D[D < 0] = 0 X[:, i] = D output = Speech().OverlapAdd2(X ** (1 / Gamma), YPhase, int(W), int(SP * W)) return output def SSBoll79_2(self, signal, fs, T1, IS=None): """ Spectral Subtraction based on Boll 79. Amplitude spectral subtraction Includes Magnitude Averaging and Residual noise Reduction :param signal: noisy signal :param fs: sampling frequency :param IS: initial silence (noise only) length in seconds (default value is .25 sec) :return output: denoise signal """ if not IS: IS = 0.25 # seconds W = int(0.025 * fs) # window length 25ms nfft = W # overlap-add method works good with this shift value SP = 0.4 # frame shift 40% (10ms) wnd = np.hamming(W) elif isinstance(IS, float): W = int(0.025 * fs) # window length 25ms nfft = W # overlap-add method works good with this shift value SP = 0.4 # frame shift 40% (10ms) wnd = np.hamming(W) # IGNORE THIS SECTION FOR COMPATIBILITY WITH ANOTHER PROGRAM FROM HERE..... if isinstance(IS, dict): W = IS['windowsize'] SP = IS['shiftsize'] / W nfft = IS['nfft'] wnd = IS['window'] if hasattr(IS, 'IS'): IS = IS['IS'] else: IS = 0.25 # .......IGNORE THIS SECTION FOR COMPATIBILITY WITH ANOTHER PROGRAM T0 HERE NIS = int((IS * fs - W) / (SP * W) + 1) # number of initial silence segments Gamma = 1 # 1: magnitude, 2: power spectrum y = self.segment(signal, W, SP, wnd) Y = np.fft.fft(y, axis=0) FreqResol, NumberofFrames = Y.shape YPhase = np.angle(Y[0: int(NumberofFrames / 2) + 1, :]) # noisy speech phase Y = np.abs(Y[0: int(NumberofFrames / 2) + 1, :]) ** Gamma # Spectrogram N = np.mean(Y[:, 0:NIS].T, axis=0).T # initial noise power spectrum mean NRM = np.zeros(N.shape) # Noise Residual Maximum (Initialization) NoiseCounter = 0 NoiseLength = 9 # smoothing factor for noise updating Beta = 0.03 fn = NumberofFrames miniL = 5 voiceseg, vosl, SF, Ef = VAD().pitch_vad1(y, fn, T1, miniL) YS = Y # Y magnitude average for i in np.arange(1, NumberofFrames - 1): YS[:, i] = (YS[:, i - 1] + YS[:, i] + YS[:, i + 1]) / 3 X = np.zeros(Y.shape) D = np.zeros(FreqResol) for i in range(NumberofFrames): # Magnitude Spectrum Distance VAD NoiseFlag, SpeechFlag, NoiseCounter, Dist = VAD().vad(Y[:, i] ** (1 / Gamma), N ** (1 / Gamma), NoiseCounter) SpeechFlag = SF[i] if SpeechFlag == 0: N = (NoiseLength * N + Y[:, i]) / (NoiseLength + 1) # update and smooth noise NRM = np.maximum(NRM, YS[:, i] - N) # update maximum noise residue X[:, i] = Beta * Y[:, i] else: D = YS[:, i] - N # spectral subtraction if i > 0 and i < NumberofFrames - 1: # residual noise reduction for j in range(len(D)): if D[j] < NRM[j]: D[j] = np.min(np.array([D[j], YS[j, i - 1] - N[j], YS[j, i + 1] - N[j]])) D[D < 0] = 0 X[:, i] = D output = Speech().OverlapAdd2(X ** (1 / Gamma), YPhase, int(W), int(SP * W)) output = output / np.max(np.abs(output)) # normalized return output def Mtmpsd_ssb(self, signal, wlen, inc, NIS, a