C++封装Essentia提取Mel特征为何慢于Python实现的Librosa？

Essentia是一款基于C开发、带有Python封装的音频与音乐分析库。理论上，C实现的Essentia应比纯Python实现的Librosa更快，但我在提取大量Mel特征数据时，运行测试demo后发现处理长音频时Essentia慢于Librosa和Audioflux（一款基于C开发、带有Python封装的音频库）。

随后我采用文档中的Mel特征提取方法，速度依然很慢，怀疑是否存在使用方式错误。以下是我的测试代码：

import time
import numpy as np
import essentia
import essentia.standard as es
import librosa
import audioflux as af

frameSize = 2048
hopSize = 512
sample_rate = 32000
data_len = frameSize + hopSize * 999

windowing = es.Windowing(type='blackmanharris62', zeroPadding=0)
spectrum = es.Spectrum(size=frameSize)
melbands = es.MelBands(inputSize=frameSize // 2 + 1, sampleRate=sample_rate,    numberBands=128,type='power',lowFrequencyBound=0, highFrequencyBound=16000, warpingFormula='slaneyMel')

use_time = 0
for _ in range(10):
    audio = np.random.randn(data_len).astype(np.float32)
    pool = essentia.Pool()
    s = time.time()
    for frame in es.FrameGenerator(audio, frameSize=frameSize, hopSize=hopSize,  startFromZero=True, lastFrameToEndOfFile=False):
        frame_spectrum = spectrum(windowing(frame))
        frame_mel = melbands(frame_spectrum)
        pool.add('mel128', frame_mel)
    mel128 = np.array(pool['mel128'])
    use_time += time.time() - s
    # print(mel128.shape)
print('essentia use time:', use_time)

测试耗时：essentia use time: 0.5148606300354004

请问是否因我的使用方式有误导致Essentia提取Mel特征速度偏慢？

内容的提问来源于stack exchange，提问作者dorian111