def tiqu(path, weidu, logenergy, energy_1):
basedir = path
for mulu in os.listdir(basedir):
input_dir = os.path.join(basedir, mulu, "wav")
#语音文件的路劲
output_dir2 = os.path.join(basedir, mulu, 'log_yuan')
#log梅尔普系数
# output_dir3 =r"C:\Users\a7825\Desktop\工作空间\语音数据\UUDB\第一次实验\打标签\第三批\C063L\mfcc"
#mfcc
muluz.mkdir(output_dir2)
for ad_file in os.listdir(input_dir):
print(input_dir + "/" + ad_file)
(fs, audio) = wav.read(input_dir + "/" + ad_file)
if energy_1 == True:
feat, energy = fbank(audio, fs, nfilt=weidu)
np.savetxt(output_dir2 + "/" + ad_file + ".csv",
feat,
delimiter=',')
if logenergy == True:
log = logfbank(audio, fs, nfilt=weidu)
np.savetxt(output_dir2 + "/" + ad_file + ".csv",
log,
delimiter=',')
def draw_logmel(wav_file, label, feature_name, logmelband_nums=[4, 5]):
(y, sr) = librosa.load(wav_file)
rate, data = read(wav_file)
plt.figure()
plt.subplot(2, 1, 1)
D = librosa.amplitude_to_db(np.abs(librosa.stft(y)), ref=np.max)
librosa.display.specshow(D, y_axis='linear', cmap="viridis")
plt.colorbar(format='%+2.0f dB')
plt.title('Linear-frequency power spectrogram')
plt.ylim([0, 8192])
plt.subplot(2, 1, 2)
feat, _ = fbank(y, samplerate=rate, nfft=2048)
logfbank_energy = np.log(feat).T
logfbank_energy = logfbank(data, samplerate=rate, nfft=2048)
colors = ["r", "g", "b"]
for i in range(len(logmelband_nums)):
logmelband_num = logmelband_nums[i]
X = np.linspace(0, len(logfbank_energy[logmelband_num]),
len(logfbank_energy[logmelband_num]))
plt.plot(X,
logfbank_energy[logmelband_num],
'o',
markersize=5,
color=colors[i],
label="logMelFreqBands[{}]".format(logmelband_num))
#quantile_value=np.quantile(logfbank_energy[logmelband_num],0.25*2)
#plt.plot(X,[quantile_value]*len(X),markersize=2,color="r",label="quartile2")
#quantile_value=np.quantile(logfbank_energy[logmelband_num],0.25*3)
#plt.plot(X,[quantile_value]*len(X),markersize=2,color="g",label="quartile3")
plt.title('logMelFrequencyBands(de)')
plt.ylabel("Filterbank")
plt.xlabel("Frame Idx")
plt.legend(loc="upper right", prop={"size": 8})
plt.savefig("/home/jialu/voice_quality_plots/v2/logMelFreqBand/" + label +
"_" + feature_name + ".png")
def dscc(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
numcep=13,
nfilt=26,
nfft=512,
lowfreq=0,
highfreq=None,
preemph=0.97,
ceplifter=22,
appendEnergy=True,
winfunc=lambda x: numpy.ones((x, ))):
feats, energies = base.fbank(signal, samplerate, winlen, winstep, nfilt,
nfft, lowfreq, highfreq, preemph, winfunc)
feats = base.delta(feats, 2) # OBTAIN DELTA
feats = boxcox(feats)
feats = numpy.log(feats)
feats = dct(feats, type=2, axis=1, norm='ortho')[:, :numcep]
feats = base.lifter(feats, ceplifter)
if appendEnergy:
feats[:, 0] = numpy.log(
energies
) # replace first cepstral coefficient with log of frame energy
feats = base.delta(feats, 2) #verify if 2 is right
return feats
def gen_fft_features(wav, step=512, nfft=[2048, 4096], n_bands=80, log=True):
features = []
# Ignoring warnings here.
# Will warn about issues calculating MEL filters when nfft = 1024.
# Causes a strange black band at band 5ish. Will ignore for now.
with warnings.catch_warnings():
warnings.simplefilter("ignore")
for fft_size in nfft:
# mel_features is of shape [T, F]
mel_features, mel_energy = fbank(wav,
nfft=fft_size,
samplerate=44100,
nfilt=n_bands,
winfunc=np.hamming,
lowfreq=27.5,
highfreq=8000.0,
winstep=512 / 44100)
if log:
features.append(np.log10(mel_features + 1e-4))
else:
features.append(mel_features)
# Reutnrs shape [Channels, Time, Frequency]
# return np.log10(np.stack(features))
return np.stack(features)
def get_features(audio_file):
"""Get features from a file"""
signal, sample_rate = sf.read(audio_file)
feat, energy = fbank(signal, sample_rate, nfilt=FLAGS.nfilt)
feat = np.log(feat)
dfeat = delta(feat, 2)
ddfeat = delta(dfeat, 2)
return np.concatenate(
[feat, dfeat, ddfeat, np.expand_dims(energy, 1)], axis=1)
def get_features(audio_file):
'''Get features from a file'''
signal, sample_rate = sf.read(tf.gfile.FastGFile(audio_file, 'rb'))
feat, energy = fbank(signal, sample_rate, nfilt=FLAGS.nfilt)
feat = np.log(feat)
dfeat = delta(feat, 2)
ddfeat = delta(dfeat, 2)
return np.concatenate([feat, dfeat, ddfeat, np.expand_dims(energy, 1)],
axis=1)
def Features(self, data, rate, dim):
spec = np.abs(np.fft.rfft(data))
freq = np.fft.rfftfreq(len(data), d=1 / dim)
a = spec / spec.sum()
meaN = (freq * a).sum()
std = np.sqrt(np.sum(a * ((freq - meaN) ** 2)))
a_cumsum = np.cumsum(a)
mediaN = freq[len(a_cumsum[a_cumsum <= 0.5])]
modE = freq[a.argmax()]
q25 = freq[len(a_cumsum[a_cumsum <= 0.25])]
q75 = freq[len(a_cumsum[a_cumsum <= 0.75])]
IQR = q75 - q25
z = a - a.mean()
w = a.std()
skewnesS = ((z ** 3).sum() / (len(spec) - 1)) / w ** 3
kurtosiS = ((z ** 4).sum() / (len(spec) - 1)) / w ** 4
m = speech.mfcc(data,rate)
f = speech.fbank(data,rate)
l = speech.logfbank(data,rate)
s = speech.ssc(data,rate)
data = pd.DataFrame(data)
desc = data.describe()
mean = desc.loc["mean"].get(0)
mad = data.mad().get(0)
sd = desc.loc["std"].get(0)
median = data.median().get(0)
minimum = desc.loc["min"].get(0)
maximum = desc.loc["max"].get(0)
Q25 = desc.loc["25%"].get(0)
Q75 = desc.loc["75%"].get(0)
interquartileR = Q75 - Q25
skewness = data.skew().get(0)
kurtosis = data.kurtosis().get(0)
result = {
"Mean": mean, "Mad": mad, "deviation": sd, "Median": median, "Min": minimum, "Max": maximum,
"interquartileR": interquartileR, "Skewness": skewness, "Q25": Q25, "Q75": Q75, "Kurtosis": kurtosis,
"mfcc_mean": np.mean(m), "mfcc_max": np.max(m), "mfcc_min": np.min(m),
"fbank_mean": np.mean(f[0]), "fbank_max": np.max(f[0]), "fbank_min": np.min(f[0]),
"energy_mean": np.mean(f[1]), "energy_max": np.max(f[1]), "energy_min": np.min(f[1]),
"lfbank_mean": np.mean(l), "lfbank_max": np.max(l), "lfbank_min": np.min(l),
"ssc_mean": np.mean(s), "ssc_max": np.max(s), "ssc_min": np.min(s),
"meaN": meaN, "deviatioN": std, "mediaN": mediaN, "modE": modE, "IQR": IQR,
"skewnesS": skewnesS, "q25": q25, "q75": q75, "kurtosiS": kurtosiS}
return result
import scipy.io.wavfile as wav
from python_speech_features.base import mfcc, fbank, logfbank
import numpy as np
import os
input_dir = r"C:\Users\a7825\Desktop\工作空间\跑代码\打标签\F1\B\wav_16"
output_dir1 = r"C:\Users\a7825\Desktop\工作空间\跑代码\打标签\F1\B\melbank"
output_dir2 = r"C:\Users\a7825\Desktop\工作空间\跑代码\打标签\F1\B\log"
if __name__ == "__main__":
for ad_file in os.listdir(input_dir):
(fs, audio) = wav.read(input_dir + "/" + ad_file)
feature_m, feature_n = fbank(audio, fs, winfunc=np.hamming)
log = logfbank(audio, fs)
np.savetxt(output_dir1 + "/" + ad_file + ".fbank.csv",
feature_m,
delimiter=',')
np.savetxt(output_dir2 + "/" + ad_file + ".log.csv",
log,
delimiter=',')
# 拼接字符串，把单引号改成双引号居然好使
# fs, audio = wav.read(r"C:\Users\a7825\Desktop\新しいフォルダー/a.wav")
# # feature_mfcc = mfcc(audio, samplerate=fs, numcep=40, nfilt=40)
# feature_m,feature_n = fbank(audio, fs)
# feature_log = logfbank(audio, fs)
# # print(feature_mel[0].shape)
# # print(feature_m)
# # print(feature_n.shape)
# np.savetxt('fbank.csv', feature_m, delimiter = ',')
# np.savetxt('energy.csv', feature_n, delimiter= ',')
def audio_features(params, img_audio, audio_path, append_name, node_list):
output_file = params['output_file']
# create pytable atom for the features
f_atom = tables.Float32Atom()
count = 1
# keep track of the nodes for which no features could be made, places
# database contains some empty audio files
invalid = []
for node in node_list:
print(f'processing file: {count}')
count += 1
# create a group for the desired feature type
audio_node = output_file.create_group(node, params['feat'])
# get the base name of the node this feature will be appended to
base_name = node._v_name.split(append_name)[1]
# get the caption file names corresponding to the image of this node
caption_files = img_audio[base_name][1]
for cap in caption_files:
# remove extension from the caption filename
base_capt = cap.split('.')[0]
# remove folder path from file names (Places/coco database)
if '/' in base_capt:
base_capt = base_capt.split('/')[-1]
if '-' in base_capt:
base_capt = base_capt.replace('-', '_')
# read audio samples
try:
input_data, fs = librosa.load(os.path.join(audio_path, cap),
sr=None)
# in the places database some of the audiofiles are empty
if len(input_data) == 0:
break
except:
# try to repair broken files, some files had a wrong header.
# In Places I found some that could not be fixed however
try:
fix_wav(os.path.join(audio_path, cap))
#input_data = read(os.path.join(audio_path, cap))
except:
# the loop will break, if no valid audio features could
# be made for this image, the entire node is deleted.
break
# set the fft size to the power of two equal to or greater than
# the window size.
window_size = int(fs * params['t_window'])
exp = 1
while True:
if np.power(2, exp) - window_size >= 0:
fft_size = np.power(2, exp)
break
else:
exp += 1
###############################################################################
# create audio features
if params['feat'] == 'raw':
# calculate the needed frame shift, premphasize and frame
# the signal
frame_shift = int(fs * params['t_shift'])
input = sigproc.preemphasis(input_data, coeff=params['alpha'])
features = sigproc.framesig(input_data,
frame_len=window_size,
frame_step=frame_shift,
winfunc=params['windowing'])
elif params['feat'] == 'freq_spectrum':
# calculate the needed frame shift, premphasize and frame
# the signal
frame_shift = int(fs * params['t_shift'])
input = sigproc.preemphasis(input_data, coeff=params['alpha'])
frames = sigproc.framesig(input,
frame_len=window_size,
frame_step=frame_shift,
winfunc=params['windowing'])
# create the power spectrum
features = sigproc.powspec(frames, fft_size)
elif params['feat'] == 'fbanks':
# create mel filterbank features
[features, energy] = base.fbank(input_data,
samplerate=fs,
winlen=params['t_window'],
winstep=params['t_shift'],
nfilt=params['nfilters'],
nfft=fft_size,
lowfreq=0,
highfreq=None,
preemph=params['alpha'],
winfunc=params['windowing'])
elif params['feat'] == 'mfcc':
# create mfcc features
features = base.mfcc(input_data,
samplerate=fs,
winlen=params['t_window'],
winstep=params['t_shift'],
numcep=params['ncep'],
nfilt=params['nfilters'],
nfft=fft_size,
lowfreq=0,
highfreq=None,
preemph=params['alpha'],
ceplifter=0,
appendEnergy=params['use_energy'],
winfunc=params['windowing'])
# apply cepstral mean variance normalisation
if params['normalise']:
features = (features - features.mean(0)) / features.std(0)
# optionally add the deltas and double deltas
if params['use_deltas']:
single_delta = base.delta(features, params['delta_n'])
double_delta = base.delta(single_delta, params['delta_n'])
features = np.concatenate(
[features, single_delta, double_delta], 1)
###############################################################################
# create new leaf node in the feature node for the current audio
# file
feature_shape = np.shape(features)[1]
f_table = output_file.create_earray(audio_node,
append_name + base_capt,
f_atom, (0, feature_shape),
expectedrows=5000)
# append new data to the tables
f_table.append(features)
if audio_node._f_list_nodes() == []:
# keep track of all the invalid nodes for which no features could
# be made
invalid.append(node._v_name)
# remove the top node including all other features if no captions
# features could be created
output_file.remove_node(node, recursive=True)
print(invalid)
print(f'There were {len(invalid)} files that could not be processed')