def test_mel_spectrogram_is_equal():
x = np.random.randn(22050)
x = np.abs(x) / np.max(np.abs(x))
mel_npy = logmelfilterbank(
x,
22050,
fft_size=1024,
hop_size=256,
win_length=None,
window="hann",
num_mels=80,
fmin=80,
fmax=7600,
eps=1e-10,
)
mel_spectrogram = MelSpectrogram(
fs=22050,
fft_size=1024,
hop_size=256,
win_length=None,
window="hann",
num_mels=80,
fmin=80,
fmax=7600,
eps=1e-10,
).to(dtype=torch.double)
mel_torch = mel_spectrogram(torch.from_numpy(x).unsqueeze(0))
np.testing.assert_array_almost_equal(
mel_npy.transpose(1, 0).astype(np.float32),
mel_torch[0].numpy().astype(np.float32),
)
def _analyze_mlfb(self, wavf):
# read wav file as float format
x, fs = sf.read(str(wavf))
self.feats["mlfb"] = logmelfilterbank(
x,
self.conf["fs"],
hop_size=self.conf["hop_size"],
win_length=self.conf["fftl"],
window="hann",
num_mels=self.conf["mlfb_dim"],
fmin=self.conf["fmin"],
fmax=self.conf["fmax"],
eps=EPS,
)
def _analyze_mlfb(self, wavf):
# read wav file as float format
x, fs = sf.read(str(wavf))
for win_type in self.windows.keys():
if win_type == "hann":
feat_name = "mlfb"
else:
feat_name = f"mlfb_{win_type}"
self.feats[feat_name] = logmelfilterbank(
x,
self.conf["fs"],
hop_size=self.conf["hop_size"],
fft_size=self.conf["fftl"],
win_length=self.conf["fftl"],
window=self.windows[win_type],
num_mels=self.conf["mlfb_dim"],
fmin=self.conf["fmin"],
fmax=self.conf["fmax"],
eps=EPS,
)