audio_reader.py 文件源码

def midi_notes_encoding(audio):
        """
        Compute frame-based midi encoding of audio
        :param audio: 1-D array of audio time series 
        """
        pitches, magnitudes = librosa.piptrack(audio)
        pitches = np.transpose(pitches)
        magnitudes = np.transpose(magnitudes)
        lc = np.zeros((pitches.shape[0], 88), dtype=np.float32)
        for i in range(pitches.shape[0]):
            # Count non-zero entries of pitches
            nz_count = len(np.nonzero(pitches[i])[0])
            # Keep a maximum of 6 detected pitches
            num_ind_to_keep = min(nz_count, 6)
            ind_of_largest_pitches = np.argpartition(
                magnitudes[i], -num_ind_to_keep)[-num_ind_to_keep:] \
                if num_ind_to_keep != 0 else []
            # Convert the largest pitches to midi notes
            overtone_limit = librosa.midi_to_hz(96)[0]
            ind_of_largest_pitches = filter(
                lambda x: pitches[i, x] <= overtone_limit,
                ind_of_largest_pitches)
            midi_notes = librosa.hz_to_midi(pitches[i, ind_of_largest_pitches])
            midi_notes = midi_notes.round()
            # Normalize magnitudes of pitches
            midi_mags = magnitudes[i, ind_of_largest_pitches] / \
                        np.linalg.norm(magnitudes[i, ind_of_largest_pitches], 1)
            np.put(lc[i], midi_notes.astype(np.int64) - [9], midi_mags)
        return lc