Compare the audio librosa library with the Mel spectrogram in the torchaudio library

Four sections describe code generation

import  torch
import  torchaudio

import  librosa
import  numpy as  np



filename = "./sample.wav"
waveform, sample_rate = torchaudio.load(filename)
a = waveform.squeeze()



n_fft = 20
win_length = 20
hop_length = 10

# -------- part 1. torch spectrogram --------------
# pytorch spectrogram
# numpy --> tensor() , b = torch.from_numpy(a)
spec = torchaudio.transforms.Spectrogram(n_fft, win_length, hop_length)
spec_pytorch = spec(torch.Tensor(a))
print(" 1. torchaudio way spectrogram: ", spec_pytorch.shape)


# ---------- part 2. librosa spectrogram --------------
# librosa spectrogram
def Spectrogram(input, n_fft, win_length, hop_length):
    stft_librosa = librosa.stft(y=input,
                            hop_length=hop_length,
                            n_fft=n_fft)
    spec_librosa = pow(np.abs(stft_librosa),2)
    return spec_librosa


a = a.numpy()
spec_librosa = Spectrogram(a, n_fft, win_length, hop_length)
print(" 2. librosa way spectrogram: ", spec_librosa.shape)

# --------------part 3. torchaudio Mel spectrogram -----------------

n_fft = 20
win_length = 20
hop_length = 10
sample_rate = 16000
mel_len = 12
mel_spec = torchaudio.transforms.MelSpectrogram(sample_rate, n_fft, win_length, hop_length, n_mels=mel_len)
mel_out = mel_spec(torch.tensor(a).to(torch.float))
print(" 3.1 torchaudio way Mel spectrogram: ",  mel_out.shape)

#  Use  torchaudio  Decomposition implementation , melspectrogram;

from torchaudio import functional as F
n_fft = 20
win_length = 20
hop_length = 10
sample_rate = 16000
mel_len = 12
stft_len = n_fft//2+1

spec = torchaudio.transforms.Spectrogram(n_fft, win_length, hop_length)
spec_out = spec(torch.Tensor(a))
#  obtain melscale The transformation matrix of 
tmp_fb = F.create_fb_matrix(stft_len, 0, sample_rate//2, mel_len, sample_rate)
mel_out1 = torch.matmul(spec_out.transpose(0,1), tmp_fb).transpose(0,1)
print(" 3.2 torchaudio way split step to generate Mel spectrogram: ",  mel_out1.shape)

# --------------part 4 librosa Mel spectrogram -----------------

from torchaudio import functional as F

n_fft = 20
win_length = 20
hop_length = 10
sample_rate = 16000
mel_len = 12
stft_len = n_fft//2+1

tmp_fb = F.create_fb_matrix(stft_len, 0, sample_rate//2, mel_len, sample_rate)

def Spectrogram(input, n_fft, win_length, hop_length):
    stft_librosa = librosa.stft(y=input,
                            hop_length=hop_length,
                            n_fft=n_fft)
    spec_librosa = pow(np.abs(stft_librosa),2)
    return spec_librosa

spec_librosa2 = Spectrogram(a, n_fft, win_length, hop_length)

mel_out2 = np.transpose(np.matmul(np.transpose( spec_librosa2),tmp_fb.numpy()))
print("4. librosa way to generate Mel spectrogram: ",  mel_out2.shape)