spk-clustering.py 文件源码

def glr(arr1, arr2):
    """Generalized Likelihood Ratio"""
    N1 = arr1.shape[0]
    N2 = arr2.shape[0]
    S1 = np.cov(arr1, rowvar=0)
    S2 = np.cov(arr2, rowvar=0)
    N = float(N1 + N2)
    # This is COV only version, not optimized (revise) but more robust
    # to environment noise conditions.
    # See Ulpu thesis pages 30-31, also Gish et al. "Segregation of
    # Speakers for Speech Recognition and Speaker Identification"
    d = -(N / 2.0) * ((N1 / N) * np.log(det(S1)) + (N2 / N) * np.log(det(S2))
                      - np.log(det((N1 / N) * S1 + (N2 / N) * S2)))
    # Ulpu version:
    # Includes the mean, theoretically less robust
    # arr = features[start:start+2*winsize]
    # S = cov(arr, rowvar=0)
    # d = -0.5*(N1*log(det(S1))+N2*log(det(S2))-N*log(det(S)))
    return d