Upload wsola.py

wsola.py

@@ -0,0 +1,67 @@
+# -*- coding: utf-8 -*-
+
+import numpy as np
+from scipy.signal import correlate2d
+from skimage.util import view_as_windows
+class WSOLA(object):
+
+    def __init__(self, fs, speech_rate, shiftms=10):
+        self.fs = fs
+        self.speech_rate = speech_rate
+
+        self.shiftms = shiftms  # Hs的时间
+        self.sl = int(self.fs * self.shiftms / 1000)  # Hs的长度
+        self.fl = self.sl * 2  # 帧长 Hs长度的两倍
+        self.epstep = int(self.sl * self.speech_rate)  # Ha的长度
+        self.win = np.hanning(self.fl)  # 窗函数
+
+    def duration_modification(self, x):
+        wlen = len(x)  # Lin
+        #Lout = Lin/rate
+        wsolaed = np.zeros(int(wlen / self.speech_rate), dtype='d') 
+        # 初始化
+        sp = self.sl * 2 # x'm的中心
+        rp = sp + self.sl # x~m的中心
+        ep = sp + self.epstep # x+(m+1)的中心
+        outp = self.sl
+        # allocate first frame of waveform to outp
+        wsolaed[:outp] = x[:outp]
+
+        while wlen > ep + self.fl:
+            # copy wavform
+            ref = x[rp - self.sl:rp + self.sl] # x~m
+            buff = x[ep - self.fl:ep + self.fl] # 搜索区域
+
+            # search minimum distance bepween ref and buff
+            delta = self._search_minimum_distance(ref, buff)
+            epd = ep + delta
+
+            # store WSOLAed waveform using over-lap add
+            spdata = x[sp:sp + self.sl] * self.win[self.sl:] # x'm的右半帧
+            epdata = x[epd - self.sl:epd] * self.win[:self.sl] # x~m的左半帧
+            if len(spdata) == len(wsolaed[outp:outp + self.sl]):
+                wsolaed[outp:outp + self.sl] = spdata + epdata
+            else:
+                wsolaed_len = len(wsolaed[outp:outp + self.sl])
+                wsolaed[outp:outp + self.sl] = spdata[:wsolaed_len] + \
+                    epdata[:wsolaed_len]
+
+            outp += self.sl
+
+            # 计算下一帧的起始位置
+            sp = epd
+            rp = sp + self.sl
+            ep += self.epstep
+
+        return wsolaed
+
+    def _search_minimum_distance(self, ref, buff):
+        if len(ref) < self.fl:
+            ref = np.r_[ref, np.zeros(self.fl - len(ref))]
+
+        # slicing and windowing one sample by one
+        buffmat = view_as_windows(buff, self.fl) * self.win
+        refwin = np.array(ref * self.win).reshape(1, self.fl)
+        corr = correlate2d(buffmat, refwin, mode='valid')
+
+        return np.argmax(corr) - self.sl