test_preprocessing.py 文件源码

def test_dtw_aligner():
    x, fs = librosa.load(example_audio_file(), sr=None)
    assert fs == 16000
    x_fast = librosa.effects.time_stretch(x, 2.0)

    X = _get_mcep(x, fs)
    Y = _get_mcep(x_fast, fs)

    D = X.shape[-1]

    # Create padded pair
    X, Y = adjast_frame_lengths(X, Y, divisible_by=2)

    # Add utterance axis
    X = X.reshape(1, -1, D)
    Y = Y.reshape(1, -1, D)

    X_aligned, Y_aligned = DTWAligner().transform((X, Y))
    assert X_aligned.shape == Y_aligned.shape
    assert np.linalg.norm(X_aligned - Y_aligned) < np.linalg.norm(X - Y)

    X_aligned, Y_aligned = IterativeDTWAligner(
        n_iter=2, max_iter_gmm=10, n_components_gmm=2).transform((X, Y))
    assert X_aligned.shape == Y_aligned.shape
    assert np.linalg.norm(X_aligned - Y_aligned) < np.linalg.norm(X - Y)

    # Custom dist function
    from nnmnkwii.metrics import melcd
    X_aligned, Y_aligned = DTWAligner(dist=melcd).transform((X, Y))
    assert np.linalg.norm(X_aligned - Y_aligned) < np.linalg.norm(X - Y)