Merge pull request #34 from peach-water/dct-numpy-reconstitution

Implement the DCT function using numpy

python/rapid_paraformer/kaldifeat/feature.py

@@ -1,5 +1,4 @@
 import numpy as np
-from scipy.fftpack import dct
 
 
 # ---------- feature-window ----------
@@ -137,6 +136,28 @@ def extract_window(waveform, blackman_coeff, dither, window_size, window_shift,
 
 # ---------- feature-functions ----------
 
+def dct_np(x: np.array, norm=None):
+    x_shape = x.shape
+    N = x_shape[-1]
+
+    v = np.hstack([x[:, ::2], x[:, 1::2][:, ::-1]])
+    Vc = np.fft.fft(v)
+    Vc = np.dstack([np.real(Vc), np.imag(Vc)])
+
+    k = - np.arange(N, dtype=x.dtype)[None, :] * np.pi / (2*N)
+    W_r = np.cos(k)
+    W_i = np.sin(k)
+
+    V = Vc[:, :, 0] * W_r - Vc[:, :, 1] * W_i
+
+    if norm == "ortho":
+        V[:, 0] /= np.sqrt(N) * 2
+        V[:, 1:] /= np.sqrt(N/2) * 2
+
+    V = 2 * V
+    return V
+
+
 def compute_spectrum(frames, n):
     complex_spec = np.fft.rfft(frames, n)
     return np.absolute(complex_spec)
@@ -424,7 +445,7 @@ def compute_mfcc_feats(
         window_type=window_type,
         dtype=dtype
     )
-    feat = dct(feat, type=2, axis=1, norm='ortho')[:, :num_ceps]
+    feat = dct_np(feat, norm="ortho")[:, :num_ceps]
     lifter_coeffs = compute_lifter_coeffs(cepstral_lifter, num_ceps).astype(dtype)
     feat = feat * lifter_coeffs
     if use_energy: