def load_audio_file(file_path: Path,
sr: int,
mono: Optional[bool] = True,
offset: Optional[float] = 0,
duration: Optional[Union[float, None]] = None) \
-> ndarray:
"""Wrapper for loading audio file.
TODO: Add handling of cases where librosa fails.
:param file_path: File path of the audio file.
:type file_path: pathlib.Path
:param sr: Sampling frequency to be used. If different\
from actual, data are resampled.
:type sr: int
:param mono: Load file as mono? Defaults to True
:type mono: bool, optional
:param offset: Offset for reading the file. Default to 0.0
:type offset: float, optional
:param duration: Duration of the returned data. Defaults to None.
:type duration: float|None, optional
:return: Audio data as numpy array of shape (channels x samples), \
if channels >= 2, else (samples, )
:rtype: numpy.ndarray
"""
return librosa_load(path=str(file_path),
sr=sr,
mono=mono,
offset=offset,
duration=duration)[0]
def __call__(self, audiopath):
audio, _ = librosa_load(audiopath, sr=self.samplerate, mono=True)
chromagram = chroma_cens(y=audio,
sr=self.samplerate,
hop_length=self.step_size)
chromagram = np.swapaxes(chromagram, 0, 1)
chromagram = np.roll(chromagram, 3, axis=1)
return chromagram, self.get_frame_times(chromagram)
def transform_playlist(wav_directory: str):
wav_list = keep_files(os_listdir(wav_directory), 'wav')
transformer = Transform()
transforms_dir = create_sub_dir(wav_directory, 'transforms')
transforms_sub_dirs = ['preciser_spectrogram', 'preciser_melspectrogram', 'preciser_cqt', 'locator_v1_cqt',
'locator_v2_spectrogram', 'locator_v2_melspectrogram']
for sub_dir in transforms_sub_dirs:
create_sub_dir(transforms_dir, sub_dir)
for wav_file in wav_list:
samples, sample_rate = librosa_load(os_path_join(wav_directory, wav_file), mono=True, sr=44100)
preciser_spectrogram = transformer.get_spectrogram(sample_rate, samples, 10, 512)
preciser_melspectrogram = transformer.get_melspectrogram(sample_rate, samples, 10, 256)
preciser_cqt = transformer.get_constant_q(sample_rate, samples, 108)
locator_v1_cqt = transformer.get_constant_q(sample_rate, samples, 108)
locator_v2_spectrogram = transformer.get_spectrogram(sample_rate, samples, 100, 4411)
locator_v2_melspectrogram = transformer.get_melspectrogram(sample_rate, samples, 100, 256)
transforms = {'preciser_spectrogram': preciser_spectrogram,
'preciser_melspectrogram': preciser_melspectrogram,
'preciser_cqt': preciser_cqt,
'locator_v1_cqt': locator_v1_cqt,
'locator_v2_spectrogram': locator_v2_spectrogram,
'locator_v2_melspectrogram': locator_v2_melspectrogram}
for transform in transforms:
imsave(
os_path_join(
transforms_dir,
transform,
wav_file + '.jpg'),
np_transpose(transforms[transform])
)
print('Saved', wav_file + '.jpg')
audio_read_flag = 0 # 1表示需要提取特征，0表示从csv读取已提取好的特征
mfcc_train_root = './mfcc_tmp/mfcc_train.npy'
mfcc_test_root = './mfcc_tmp/mfcc_test.npy'
if not os_path_exists('./mfcc_tmp/'):
makedirs('./mfcc_tmp/')
audio_read_flag = 1
audio_train_data = np.empty([train_pos_num + train_neg_num, 2 * parameter_num])
audio_test_data = np.empty([test_num, 2 * parameter_num])
if audio_read_flag == 1:
audio_data1 = np.empty([train_pos_num, 2 * parameter_num])
audio_data2 = np.empty([train_neg_num, 2 * parameter_num])
for i in range(0, train_pos_num):
y, sr = librosa_load(dataset_root + '/train/positive/' + str(i) +
'/audio.wav',
sr=16000) # initial sr=16000
mfccs = mfcc(y=y, sr=sr, n_mfcc=parameter_num)
for j in range(0, parameter_num):
audio_data1[i][j] = np.mean(mfccs[j][:])
audio_data1[i][j + parameter_num] = np.var(mfccs[j][:])
print("Audio positive read finished")
for i in range(0, train_neg_num):
y, sr = librosa_load(dataset_root + '/train/negative/' + str(i) +
'/audio.wav',
sr=16000)
mfccs = mfcc(y=y, sr=sr, n_mfcc=parameter_num)
for j in range(0, parameter_num):
audio_data2[i][j] = np.mean(mfccs[j][:])
audio_data2[i][j + parameter_num] = np.var(mfccs[j][:])
def check_file(in_path=""):
try:
librosa_load(path=in_path, sr=_sr)
except Exception as e:
remove(in_path)
pp("Delete: " + basename(in_path))
def _load_file_raw(self, pkl):
self.y, self.sr = librosa_load(path=self.input_file, sr=self.sr, duration=self.load_duration)
if Values.USE_PKL:
pickle_dump([self.y, self.sr], open(pkl, "wb"))