def build_audio3(is_clean, params, spk_index, audio_samples_length=-1):
'''Construct an audio signal from source files of primary speaker
Returns a list of all segments belonging to primary speaker'''
fs_output = params['fs']
silence_length = params['silence_length']
if audio_samples_length == -1:
audio_samples_length = int(params['audio_length'] * params['fs'])
output_audio = np.zeros(0)
remaining_length = audio_samples_length
files_used = []
clipped_files = []
if is_clean:
source_files = params['cleanfilenames']
idx = spk_index
input_audio, fs_input = audioread(source_files[idx])
if fs_input != fs_output:
input_audio = librosa.resample(input_audio, fs_input, fs_output)
num_segs = int(input_audio.shape[0] /
(params['fs'] * params['audio_length']))
# if current file is longer than remaining desired length, and this is
# noise generation or this is training set, subsample it randomly
remaining_length = input_audio.shape[0]
start_idx = 0
audios_all = []
for seg in range(0, num_segs):
# print(seg)
audio_temp = input_audio[seg * audio_samples_length:(seg + 1) *
audio_samples_length]
if not is_clipped(audio_temp):
audios_all.append(audio_temp)
files_used.append(source_files[idx])
else:
clipped_files.append(source_files[idx])
return audios_all, files_used, clipped_files
def build_audio(is_clean, params, index, audio_samples_length=-1):
'''Construct an audio signal from source files'''
fs_output = params['fs']
silence_length = params['silence_length']
if audio_samples_length == -1:
audio_samples_length = int(params['audio_length']*params['fs'])
output_audio = np.zeros(0)
remaining_length = audio_samples_length
files_used = []
clipped_files = []
if is_clean:
source_files = params['cleanfilenames']
idx = index
else:
if 'noisefilenames' in params.keys():
source_files = params['noisefilenames']
idx = index
# if noise files are organized into individual subdirectories, pick a directory randomly
else:
noisedirs = params['noisedirs']
# pick a noise category randomly
idx_n_dir = np.random.randint(0, np.size(noisedirs))
source_files = glob.glob(os.path.join(noisedirs[idx_n_dir],
params['audioformat']))
shuffle(source_files)
# pick a noise source file index randomly
idx = np.random.randint(0, np.size(source_files))
# initialize silence
silence = np.zeros(int(fs_output*silence_length))
# iterate through multiple clips until we have a long enough signal
tries_left = MAXTRIES
while remaining_length > 0 and tries_left > 0:
# read next audio file and resample if necessary
idx = (idx + 1) % np.size(source_files)
input_audio, fs_input = audioread(source_files[idx])
if input_audio is None:
sys.stderr.write("WARNING: Cannot read file: %s\n" % source_files[idx])
continue
if fs_input != fs_output:
input_audio = librosa.resample(input_audio, fs_input, fs_output)
# if current file is longer than remaining desired length, and this is
# noise generation or this is training set, subsample it randomly
if len(input_audio) > remaining_length and (not is_clean or not params['is_test_set']):
idx_seg = np.random.randint(0, len(input_audio)-remaining_length)
input_audio = input_audio[idx_seg:idx_seg+remaining_length]
# check for clipping, and if found move onto next file
if is_clipped(input_audio):
clipped_files.append(source_files[idx])
tries_left -= 1
continue
# concatenate current input audio to output audio stream
files_used.append(source_files[idx])
output_audio = np.append(output_audio, input_audio)
remaining_length -= len(input_audio)
# add some silence if we have not reached desired audio length
if remaining_length > 0:
silence_len = min(remaining_length, len(silence))
output_audio = np.append(output_audio, silence[:silence_len])
remaining_length -= silence_len
if tries_left == 0 and not is_clean and 'noisedirs' in params.keys():
print("There are not enough non-clipped files in the " + noisedirs[idx_n_dir] + \
" directory to complete the audio build")
return [], [], clipped_files, idx
return output_audio, files_used, clipped_files, idx
def main_gen(params):
'''Calls gen_audio() to generate the audio signals, verifies that they meet
the requirements, and writes the files to storage'''
clean_source_files = []
clean_clipped_files = []
clean_low_activity_files = []
noise_source_files = []
noise_clipped_files = []
noise_low_activity_files = []
clean_index = 0
noise_index = 0
file_num = params['fileindex_start']
while file_num <= params['fileindex_end']:
# generate clean speech
clean, clean_sf, clean_cf, clean_laf, clean_index = \
gen_audio(True, params, clean_index)
# add reverb with selected RIR
rir_index = random.randint(0,len(params['myrir'])-1)
my_rir = os.path.normpath(os.path.join('datasets', 'impulse_responses', params['myrir'][rir_index]))
(fs_rir,samples_rir) = wavfile.read(my_rir)
my_channel = int(params['mychannel'][rir_index])
if samples_rir.ndim==1:
samples_rir_ch = np.array(samples_rir)
elif my_channel > 1:
samples_rir_ch = samples_rir[:, my_channel -1]
else:
samples_rir_ch = samples_rir[:, my_channel -1]
#print(samples_rir.shape)
#print(my_channel)
clean = add_pyreverb(clean, samples_rir_ch)
# generate noise
noise, noise_sf, noise_cf, noise_laf, noise_index = \
gen_audio(False, params, noise_index, len(clean))
clean_clipped_files += clean_cf
clean_low_activity_files += clean_laf
noise_clipped_files += noise_cf
noise_low_activity_files += noise_laf
# get rir files and config
# mix clean speech and noise
# if specified, use specified SNR value
if not params['randomize_snr']:
snr = params['snr']
# use a randomly sampled SNR value between the specified bounds
else:
snr = np.random.randint(params['snr_lower'], params['snr_upper'])
clean_snr, noise_snr, noisy_snr, target_level = segmental_snr_mixer(params=params,
clean=clean,
noise=noise,
snr=snr)
# Uncomment the below lines if you need segmental SNR and comment the above lines using snr_mixer
#clean_snr, noise_snr, noisy_snr, target_level = segmental_snr_mixer(params=params,
# clean=clean,
# noise=noise,
# snr=snr)
# unexpected clipping
if is_clipped(clean_snr) or is_clipped(noise_snr) or is_clipped(noisy_snr):
print("Warning: File #" + str(file_num) + " has unexpected clipping, " + \
"returning without writing audio to disk")
continue
clean_source_files += clean_sf
noise_source_files += noise_sf
# write resultant audio streams to files
hyphen = '-'
clean_source_filenamesonly = [i[:-4].split(os.path.sep)[-1] for i in clean_sf]
clean_files_joined = hyphen.join(clean_source_filenamesonly)[:MAXFILELEN]
noise_source_filenamesonly = [i[:-4].split(os.path.sep)[-1] for i in noise_sf]
noise_files_joined = hyphen.join(noise_source_filenamesonly)[:MAXFILELEN]
noisyfilename = clean_files_joined + '_' + noise_files_joined + '_snr' + \
str(snr) + '_tl' + str(target_level) + '_fileid_' + str(file_num) + '.wav'
cleanfilename = 'clean_fileid_'+str(file_num)+'.wav'
noisefilename = 'noise_fileid_'+str(file_num)+'.wav'
noisypath = os.path.join(params['noisyspeech_dir'], noisyfilename)
cleanpath = os.path.join(params['clean_proc_dir'], cleanfilename)
noisepath = os.path.join(params['noise_proc_dir'], noisefilename)
audio_signals = [noisy_snr, clean_snr, noise_snr]
file_paths = [noisypath, cleanpath, noisepath]
file_num += 1
for i in range(len(audio_signals)):
try:
audiowrite(file_paths[i], audio_signals[i], params['fs'])
except Exception as e:
print(str(e))
return clean_source_files, clean_clipped_files, clean_low_activity_files, \
noise_source_files, noise_clipped_files, noise_low_activity_files
def build_audio(is_clean, params, filenum, audio_samples_length=-1):
'''Construct an audio signal from source files'''
fs_output = params['fs']
silence_length = params['silence_length']
if audio_samples_length == -1:
audio_samples_length = int(params['audio_length']*params['fs'])
output_audio = np.zeros(0)
remaining_length = audio_samples_length
files_used = []
clipped_files = []
global clean_counter, noise_counter
if is_clean:
source_files = params['cleanfilenames']
idx_counter = clean_counter
else:
source_files = params['noisefilenames']
idx_counter = noise_counter
# initialize silence
silence = np.zeros(int(fs_output*silence_length))
# iterate through multiple clips until we have a long enough signal
tries_left = MAXTRIES
while remaining_length > 0 and tries_left > 0:
# read next audio file and resample if necessary
with idx_counter.get_lock():
idx_counter.value += 1
idx = idx_counter.value % np.size(source_files)
input_audio, fs_input = audioread(source_files[idx])
if fs_input != fs_output:
input_audio = librosa.resample(input_audio, fs_input, fs_output)
# if current file is longer than remaining desired length, and this is
# noise generation or this is training set, subsample it randomly
if len(input_audio) > remaining_length and (not is_clean or not params['is_test_set']):
idx_seg = np.random.randint(0, len(input_audio)-remaining_length)
input_audio = input_audio[idx_seg:idx_seg+remaining_length]
# check for clipping, and if found move onto next file
if is_clipped(input_audio):
clipped_files.append(source_files[idx])
tries_left -= 1
continue
# concatenate current input audio to output audio stream
files_used.append(source_files[idx])
output_audio = np.append(output_audio, input_audio)
remaining_length -= len(input_audio)
# add some silence if we have not reached desired audio length
if remaining_length > 0:
silence_len = min(remaining_length, len(silence))
output_audio = np.append(output_audio, silence[:silence_len])
remaining_length -= silence_len
if tries_left == 0:
print("Audio generation failed for filenum " + str(filenum))
return [], [], clipped_files
return output_audio, files_used, clipped_files
def main_gen(params, filenum):
'''Calls gen_audio() to generate the audio signals, verifies that they meet
the requirements, and writes the files to storage'''
print("Generating file #" + str(filenum))
clean_clipped_files = []
clean_low_activity_files = []
noise_clipped_files = []
noise_low_activity_files = []
while True:
# generate clean speech
clean, clean_source_files, clean_cf, clean_laf = \
gen_audio(True, params, filenum)
# generate noise
noise, noise_source_files, noise_cf, noise_laf = \
gen_audio(False, params, filenum, len(clean))
clean_clipped_files += clean_cf
clean_low_activity_files += clean_laf
noise_clipped_files += noise_cf
noise_low_activity_files += noise_laf
# mix clean speech and noise
# if specified, use specified SNR value
if not params['randomize_snr']:
snr = params['snr']
# use a randomly sampled SNR value between the specified bounds
else:
snr = np.random.randint(params['snr_lower'], params['snr_upper'])
clean_snr, noise_snr, noisy_snr, target_level = snr_mixer(params=params,
clean=clean,
noise=noise,
snr=snr)
# Uncomment the below lines if you need segmental SNR and comment the above lines using snr_mixer
#clean_snr, noise_snr, noisy_snr, target_level = segmental_snr_mixer(params=params,
# clean=clean,
# noise=noise,
# snr=snr)
# unexpected clipping
if is_clipped(clean_snr) or is_clipped(noise_snr) or is_clipped(noisy_snr):
continue
else:
break
# write resultant audio streams to files
hyphen = '-'
clean_source_filenamesonly = [i[:-4].split(os.path.sep)[-1] for i in clean_source_files]
clean_files_joined = hyphen.join(clean_source_filenamesonly)[:MAXFILELEN]
noise_source_filenamesonly = [i[:-4].split(os.path.sep)[-1] for i in noise_source_files]
noise_files_joined = hyphen.join(noise_source_filenamesonly)[:MAXFILELEN]
noisyfilename = clean_files_joined + '_' + noise_files_joined + '_snr' + \
str(snr) + '_fileid_' + str(filenum) + '.wav'
cleanfilename = 'clean_fileid_'+str(filenum)+'.wav'
noisefilename = 'noise_fileid_'+str(filenum)+'.wav'
noisypath = os.path.join(params['noisyspeech_dir'], noisyfilename)
cleanpath = os.path.join(params['clean_proc_dir'], cleanfilename)
noisepath = os.path.join(params['noise_proc_dir'], noisefilename)
audio_signals = [noisy_snr, clean_snr, noise_snr]
file_paths = [noisypath, cleanpath, noisepath]
for i in range(len(audio_signals)):
try:
audiowrite(file_paths[i], audio_signals[i], params['fs'])
except Exception as e:
print(str(e))
pass
return clean_source_files, clean_clipped_files, clean_low_activity_files, \
noise_source_files, noise_clipped_files, noise_low_activity_files
def main_gen(params):
'''Calls gen_audio() to generate the audio signals, verifies that they meet
the requirements, and writes the files to storage'''
clean_source_files = []
clean_clipped_files = []
clean_low_activity_files = []
clean_source_files2 = []
clean_clipped_files2 = []
clean_low_activity_files2 = []
noise_source_files = []
noise_clipped_files = []
noise_low_activity_files = []
clean_index = 0
clean_index2 = 0
noise_index = 0
file_num = params['fileindex_start']
cleanfilenames = params['cleanfilenames']
cleanfilenames2 = params['cleanfilenames2']
# spk_index = 0 # start of speaker index
num_spk = len(cleanfilenames)
while file_num <= params['fileindex_end']:
# generate clean speech
#clean, clean_sf, clean_cf, clean_laf, clean_index = \
# gen_audio(True, params, clean_index)
spk_index = random.randint(0, len(params['cleanfilenames']) - 1)
chosen_clean = gen_audio3(True, params, spk_index)
num_clips = int(len(chosen_clean))
#(True, params, clean_index)
# add reverb with selected RIR
#rir_index = random.randint(0,len(params['myrir'])-1)
num_to_select1 = 1
rirfilenames = params['myrir']
chosen_clean_reverb = []
for clean in chosen_clean:
myrir = random.sample(rirfilenames, num_to_select1)
(fs_rir, samples_rir) = wavfile.read(myrir[0])
if len(samples_rir.shape) > 1:
channel = random.randint(0, len(samples_rir.shape) - 1)
samples_rir_ch = samples_rir[:, channel]
else:
samples_rir_ch = samples_rir
clean_reverb = add_pyreverb(clean, samples_rir_ch)
chosen_clean_reverb.append(clean_reverb)
# add secondary speech and/or noise for each chunck of primary speech
for chose_primary in chosen_clean_reverb:
index2 = random.randint(0, len(params['cleanfilenames2']) - 1)
clean2, clean_sf, clean_cf, clean_laf, clean_index = \
gen_audio2(True, params, index2, chose_primary.shape[0])
noise_index = random.randint(0, len(params['noisefilenames']) - 1)
# generate noise
noise, noise_sf, noise_cf, noise_laf, noise_index = \
gen_audio(False, params, noise_index, chose_primary.shape[0])
# if specified, use specified SNR value
if not params['randomize_snr']:
snr = params['snr']
# use a randomly sampled SNR value between the specified bounds
else:
snr = np.random.randint(params['snr_lower'],
params['snr_upper'])
snr2 = np.random.randint(params['snr_lower'],
params['snr_upper'])
snr3 = np.random.randint(params['snr_lower'],
params['snr_upper'])
# 1. Primary(clean) + Noise
clean_snr, noise_snr, noisy_snr, target_level = segmental_snr_mixer(
params=params, clean=chose_primary, noise=noise, snr=snr)
# 2. Primary + Secondary
clean_snr2, noise_snr2, noisy_snr2, target_level2 = segmental_snr_mixer(
params=params, clean=chose_primary, noise=clean2, snr=snr2)
# 3. Primary + Seconday (clean2) + Noise
clean_snr3, noise_snr3, noisy_snr3, target_level3 = segmental_snr_mixer(
params=params, clean=noisy_snr2, noise=noise, snr=snr3)
# unexpected clipping
#if is_clipped(clean_snr) or is_clipped(noise_snr2) or is_clipped(noisy_snr2):
if is_clipped(clean_snr) or is_clipped(noisy_snr):
print("Warning: File #" + str(file_num) + " has unexpected clipping, " + \
"returning without writing audio to disk")
continue
if is_clipped(clean_snr2) or is_clipped(noisy_snr2):
print("Warning: File #" + str(file_num) + " has unexpected clipping, " + \
"returning without writing audio to disk")
continue
if is_clipped(clean_snr3) or is_clipped(noisy_snr3):
print("Warning: File #" + str(file_num) + " has unexpected clipping, " + \
"returning without writing audio to disk")
continue
clean_source_files += clean_sf
noise_source_files += noise_sf
# write resultant audio streams to files
hyphen = '-'
clean_source_filenamesonly = [
i[:-4].split(os.path.sep)[-1] for i in clean_sf
]
clean_files_joined = hyphen.join(
clean_source_filenamesonly)[:MAXFILELEN]
noise_source_filenamesonly = [
i[:-4].split(os.path.sep)[-1] for i in noise_sf
]
noise_files_joined = hyphen.join(
noise_source_filenamesonly)[:MAXFILELEN]
noisyfilename = 'primary_noisy_fileid_' + str(file_num) + '_' + clean_files_joined + '_' + noise_files_joined + '_snr' + \
str(snr) + '_tl' + str(target_level) + '.wav'
cleanfilename = 'clean_fileid_' + str(file_num) + '.wav'
noisefilename = 'noise_fileid_' + str(file_num) + '.wav'
noisypath = os.path.join(params['noisyspeech_dir'], noisyfilename)
cleanpath = os.path.join(params['clean_proc_dir'], cleanfilename)
noisepath = os.path.join(params['noise_proc_dir'], noisefilename)
noisyfilename2 = 'ps_noisy_fileid_'+ str(file_num) + '_' + clean_files_joined + '_' + noise_files_joined + '_snr' + \
str(snr) + '_tl' + str(target_level) + '.wav'
cleanfilename2 = 'ps_clean_fileid_' + str(file_num) + '.wav'
noisefilename2 = 'ps_noise_fileid_' + str(file_num) + '.wav'
noisypath2 = os.path.join(params['noisyspeech_dir'],
noisyfilename2)
cleanpath2 = os.path.join(params['clean_proc_dir'], cleanfilename2)
noisepath2 = os.path.join(params['noise_proc_dir'], noisefilename2)
noisyfilename3 = 'psn_noisy_fileid_' + str(file_num) + '_' +clean_files_joined + '_' + noise_files_joined + '_snr' + \
str(snr) + '_tl' + str(target_level) + '.wav'
cleanfilename3 = 'psn_clean_fileid_' + str(file_num) + '.wav'
noisefilename3 = 'psn_noise_fileid_' + str(file_num) + '.wav'
noisypath3 = os.path.join(params['noisyspeech_dir'],
noisyfilename3)
cleanpath3 = os.path.join(params['clean_proc_dir'], cleanfilename3)
noisepath3 = os.path.join(params['noise_proc_dir'], noisefilename3)
audio_signals = [noisy_snr, clean_snr, noise_snr]
file_paths = [noisypath, cleanpath, noisepath]
audio_signals2 = [noisy_snr2, clean_snr, noise_snr2]
file_paths2 = [noisypath2, cleanpath2, noisepath2]
audio_signals3 = [noisy_snr3, clean_snr, noise_snr3]
file_paths3 = [noisypath3, cleanpath3, noisepath3]
file_num += 1 # file_num = file_num + 3*num_clips
for i in range(len(audio_signals)):
try:
audiowrite(file_paths[i], audio_signals[i], params['fs'])
audiowrite(file_paths2[i], audio_signals2[i], params['fs'])
audiowrite(file_paths3[i], audio_signals3[i], params['fs'])
except Exception as e:
print(str(e))
# for i in range(len(audio_signals2)):
# try:
# except Exception as e:
# print(str(e))
# for i in range(len(audio_signals3)):
# try:
# audiowrite(file_paths3[i], audio_signals3[i], params['fs'])
# except Exception as e:
# print(str(e))
return clean_source_files, clean_clipped_files, clean_low_activity_files, \
noise_source_files, noise_clipped_files, noise_low_activity_files