def separate(ckpt, mtgph, input_path, output_path):
model = inference.SeparationModel(ckpt, mtgph)
if not os.path.exists(input_path):
raise Exception("Wrong input file {}".format(input_path))
file_list = [input_path]
from scipy.io.wavfile import read as read_wav
from scipy.io.wavfile import write as write_wav
sr, f = read_wav(file_list[0])
with model.graph.as_default():
dataset = data_io.wavs_to_dataset(file_list,
batch_size=1,
num_samples=len(f),
repeat=False)
# Strip batch and mic dimensions.
dataset['receiver_audio'] = dataset['receiver_audio'][0, 0]
dataset['source_images'] = dataset['source_images'][0, :, 0]
waveforms = model.sess.run(dataset)
separated_waveforms = model.separate(waveforms['receiver_audio'])
# print(separated_waveforms)
# print(separated_waveforms.shape)
if not os.path.exists(output_path):
os.makedirs(output_path)
for i in range(separated_waveforms.shape[0]):
write_wav(output_path + '/sub_target{}.wav'.format(i), sr,
separated_waveforms[i, :])
def __call__(self, feats, hparams, name=None):
linear_feats = self._logmel_to_linear(feats)
samples = self._griffin_lim(linear_feats, hparams.gl_iters)
#samples = samples / 32768
if not os.path.exists(hparams.output_directory):
os.makedirs(hparams.output_directory)
output_path = os.path.join(hparams.output_directory,
'%s.wav' % str(name))
write_wav(output_path, self.sample_rate,
(samples * np.iinfo(np.int16).max).astype(np.int16))
def save_to_wav(self, file_name):
""" Save this time series to a wav format audio file.
Parameters
----------
file_name : string
The output file name
"""
scaled = _numpy.int16(self.numpy()/max(abs(self)) * 32767)
write_wav(file_name, self.sample_rate, scaled)
def save_to_wav(self, file_name):
""" Save this time series to a wav format audio file.
Parameters
----------
file_name : string
The output file name
"""
scaled = _numpy.int16(self.numpy() / max(abs(self)) * 32767)
write_wav(file_name, int(self.sample_rate), scaled)
def wav2guitar_distortion(inFile,outFile=None, deepValue=0.5):
""" преобразовать звук в файле таким образом, как будто он
прошел через дисторшн примочку"""
if outFile is None:
outFile = changeFileExt(inFile,'_distortion.wav')
fps,data_in = read_wav(inFile)
Amax = int(deepValue*np.max(data_in))
Amin = int(deepValue*np.min(data_in))
data_out = np.clip(data_in,Amin,Amax) # clip sound amplitude
write_wav(outFile,fps,data_out)
def save(self, path):
"""save Audio to .wav file using scipy.io.wavfile
Args:
self
path: destination for wav file """
if path.split(".")[-1] != "wav":
raise ValueError("file extension must be .wav")
from scipy.io.wavfile import write as write_wav
write_wav(path, self.sample_rate, self.samples)
def handleConvolve(self):
"""Runs the convolution when the convolve button is clicked.
"""
def _getFile(widget, name):
filename = widget.text()
if filename == '':
raise Exception(('Cannot run convolution: must specify the {}'
).format(name))
if not os.path.exists:
raise Exception(
('Cannot run convolution: path specified {} does not exist'
).format(name))
rate, data = read_wav(filename)
data = pcm2float(data, 'float32')
if len(data.shape) == 1:
data = data.reshape((len(data), 1))
return data, rate
try:
signal, rate = _getFile(self.signal_loc, 'signal')
filter, _ = _getFile(self.filter_loc, 'filter')
print('Signal has {} channel(s).'.format(signal.shape[1]))
print('Filter has {} channel(s).'.format(signal.shape[1]))
rs = None
count = 0
for i in range(signal.shape[1]):
for j in range(filter.shape[1]):
convolution = float2pcm(
convolve(signal[:, i], filter[:, j]), 'int16')
convolution = convolution.reshape((len(convolution), 1))
if rs is None:
rs = convolution
else:
rs = np.append(rs, convolution, axis=1)
count += 1
filename = self.saveFileDialog()
if not filename:
raise Exception('Error: Save file not specified')
write_wav(filename, rate, convolution)
self.save_loc = filename
except Exception as e:
self.output.setText(str(e))
def save_sample(model_name, model, waveglow_path, tacotron2_path, phrase_path,
filepath, sampling_rate, fp16):
if phrase_path is None:
return
phrase = torch.load(phrase_path, map_location='cpu')
if model_name == 'Tacotron2':
if waveglow_path is None:
raise Exception(
"WaveGlow checkpoint path is missing, could not generate sample"
)
with torch.no_grad():
checkpoint = torch.load(waveglow_path, map_location='cpu')
waveglow = models.get_model('WaveGlow',
checkpoint['config'],
to_fp16=False,
to_cuda=False)
waveglow.eval()
model.eval()
mel = model.infer(phrase.cuda())[0].cpu()
model.train()
if fp16:
mel = mel.float()
audio = waveglow.infer(mel, sigma=0.6)
elif model_name == 'WaveGlow':
if tacotron2_path is None:
raise Exception(
"Tacotron2 checkpoint path is missing, could not generate sample"
)
with torch.no_grad():
checkpoint = torch.load(tacotron2_path, map_location='cpu')
tacotron2 = models.get_model('Tacotron2',
checkpoint['config'],
to_fp16=False,
to_cuda=False)
tacotron2.eval()
mel = tacotron2.infer(phrase)[0].cuda()
model.eval()
if fp16:
mel = mel.half()
audio = model.infer(mel, sigma=0.6).cpu()
model.train()
if fp16:
audio = audio.float()
else:
raise NotImplementedError(
"unknown model requested: {}".format(model_name))
audio = audio[0].numpy()
audio = audio.astype('int16')
write_wav(filepath, sampling_rate, audio)
def write_wav_test(fileName):
""" записать тест для wav файла"""
fps = 44100 # cd quality
time = np.linspace(0,1,fps) # время 1 секунда
frame_raw = np.ndarray(shape=(fps),dtype = np.int16)
Ampl = 1000
notes_track = ['c','d','e','f','g','a','b']
audio_frames = []
for i in range(-2,2):
for n in notes_track:
f = note2freq(n,i)
frame_raw[:] = Ampl*np.sin(2*np.pi*f*time)
audio_frames.append(np.copy(frame_raw))
write_wav(fileName,fps,np.block(audio_frames))
def write_to_file(arr, ext, gzip=False):
global file_index, record_name
sys.stdout.write('start write to file ' + str(len(arr)) + ' values...')
sys.stdout.flush()
data_dir = 'data-temp/'
# fileprefix = 'fio-disease-'
fileprefix = record_name + '-'
if not os.path.exists(data_dir):
os.makedirs(data_dir)
filename = data_dir + fileprefix + str(file_index) + '.' + ext
if ext == 'dat':
with open(filename, 'w') as f:
arr.tofile(f)
elif ext == 'txt':
np.savetxt(filename, arr)
elif ext == 'wav':
rate = int(arr[1])
arr = arr[2:] # del record_time and rate
# scaled = np.int16(arr / np.max(np.abs(arr)) * 32767)
# write_wav(filename, rate, scaled)
write_wav(filename, rate, arr)
else:
print('wrong file extension')
file_index += 1
filesize = os.stat(filename).st_size
print(" done (", filesize, ' bytes)', sep='')
print(filename)
if gzip:
sys.stdout.write('gzip data compression: ' + str(filesize / 1000000) + 'MB...')
sys.stdout.flush()
with open(filename, 'rb') as f_in, gzip.open(filename + '.gz', 'wb') as f_out:
shutil.copyfileobj(f_in, f_out)
gzfilesize = os.stat(filename + '.gz').st_size
print(' done. File reduced to ', gzfilesize / 1000000, 'MB (%0.0f' % (gzfilesize/filesize*100), '% of uncompressed)', sep='')
def compare(
model: keras.Model,
generator: SoundGenerator,
parameters: ParameterSet,
orig_file: str,
output_dir: str,
orig_params,
length: float,
sample_rate: int,
extra: dict = {},
):
# (copy original file if given)
base_filename = orig_file.replace(".wav", "")
base_filename = re.sub(r".*/", "", base_filename)
copy_file: str = f"{output_dir}/{base_filename}_copy.wav"
regen_file: str = f"{output_dir}/{base_filename}_duplicate.wav"
reconstruct_file: str = f"{output_dir}/{base_filename}_reconstruct.wav"
print(f"Creating copy as {copy_file}")
# Load the wave file
fs, data = wavfile.read(orig_file)
# Copy original file to make sure
write_wav(copy_file, sample_rate, data)
# Decode original params, and regenerate output (make sure its correct)
orig = parameters.encoding_to_settings(orig_params)
generator.generate(orig, regen_file, length, sample_rate, extra)
# Run the wavefile into the model for prediction
X = [data]
Xd = np.expand_dims(np.vstack(X), axis=2)
# Get encoded parameters out of model
result = model.predict(Xd)[0]
# Decode prediction, and reconstruct output
predicted = parameters.encoding_to_settings(result)
generator.generate(predicted, reconstruct_file, length, sample_rate, extra)
batch_size=batch_size,
epochs=epochs,
callbacks=[tensorboard_callback])
else:
wgan.fit(train_ds, batch_size=batch_size, epochs=epochs)
# Saving the models
if args.save_model:
os.mkdir('trained')
generator.save_weights('trained/g.h5')
discriminator.save_weights('trained/d.h5')
# Generating a Sample
if args.generate:
if not os.path.exists('trained'):
print('unable to generate samples. No trained model exists')
else:
print("Generating a random Sample of audio as 'output.wav'")
if not args.train:
generator = Generator()
discriminator = Discrimiator()
try:
generator.load_weights('trained/g.h5')
discriminator.load_weights('trained/d.h5')
except:
print("An Error occured while loading the model")
noise = tf.random.normal([1, 100])
generated = generator(noise, training=False)
data = generated[0, :, :, :].reshape(128 * 128 * 1)
write_wav('output.wav', 16000, data)
def write_file(self, data: np.ndarray, filename: str, sample_rate: int):
# REVIEW: is this needed?
# int_data = (data * np.iinfo(np.int16).max).astype(int)
write_wav(filename, sample_rate, data)
def split_ltts(do_loud: bool) -> None:
"""MAKEDOC: what is split_ltts doing?"""
logg = logging.getLogger(f"c.{__name__}.split_ltts")
# logg.setLevel("INFO")
logg.debug("Start split_ltts")
# the location of this file
this_file_folder = Path(__file__).parent.absolute()
logg.debug(f"this_file_folder: {this_file_folder}")
####################################################################################
# Build the dict of sentences without training words inside #
####################################################################################
# the location of the original dataset
ltts_base_folder = Path.home() / "audiodatasets" / "LibriTTS" / "dev-clean"
logg.debug(f"ltts_base_folder: {ltts_base_folder}")
# build good_IDs dict
# from wav_ID to orig_wav_path
good_IDs: ty.Dict[str, Path] = {}
# get the words
train_words = words_types["all"]
# train_words = words_types["num"]
# count good sentences
good_sent = 0
skip_sent = 0
for reader_ID_path in ltts_base_folder.iterdir():
for chapter_ID_path in reader_ID_path.iterdir():
for file_path in chapter_ID_path.iterdir():
# extract the name of the file
file_name = file_path.name
# only process normalized files
if "normalized" not in file_name:
continue
# logg.debug(f"file_path: {file_path}")
# read the normalized transcription
norm_tra = file_path.read_text()
# skip this row if one of the training words is in the sentence
found_train_word = False
for word in train_words:
if word in norm_tra:
# logg.debug(f"Found '{word}' in {norm_tra}")
found_train_word = True
skip_sent += 1
break
if found_train_word:
continue
good_sent += 1
# build the wav path
# file_path /path/to/file/wav_ID.normalized.txt
# with stem extract wav_ID.normalized
# then remove .normalized
wav_ID = file_path.stem[:-11]
# logg.debug(f"wav_ID: {wav_ID}")
orig_wav_path = chapter_ID_path / f"{wav_ID}.wav"
# save the file path
good_IDs[wav_ID] = orig_wav_path
logg.debug(f"good_sent: {good_sent}")
logg.debug(f"skip_sent: {skip_sent}")
####################################################################################
# Create the samples #
####################################################################################
# where to save the results
ltts_proc_folder = this_file_folder / "data_raw"
if do_loud:
loud_tag = "_loud"
else:
loud_tag = ""
ltts_label = f"_other_ltts{loud_tag}"
ltts_label_folder = ltts_proc_folder / ltts_label
logg.debug(f"ltts_label_folder: {ltts_label_folder}")
if not ltts_label_folder.exists():
ltts_label_folder.mkdir(parents=True, exist_ok=True)
# the file to list the file names for the testing fold
val_list_path = ltts_proc_folder / f"validation_list_lttspeech{loud_tag}.txt"
word_val_list: ty.List[str] = []
test_list_path = ltts_proc_folder / f"testing_list_lttspeech{loud_tag}.txt"
word_test_list: ty.List[str] = []
# the sizes of the test and validation folds
val_fold_size = 0.1
test_fold_size = 0.1
# the target sr
new_sr = 16000
# how many samples to generate
max_num_samples = 5000
saved_samples = 0
# a random number generator to use
rng = np.random.default_rng(12345)
for wav_ID in good_IDs:
orig_wav_path = good_IDs[wav_ID]
logg.debug(f"good_IDs[{wav_ID}]: {good_IDs[wav_ID]}")
# load the original signal
orig_sig, orig_sr = librosa.load(orig_wav_path, sr=None)
# resample it to 16000 Hz
new_sig = librosa.resample(orig_sig, orig_sr, new_sr)
# how long is the sample to extract
if do_loud:
sample_len = new_sr // 2
else:
sample_len = new_sr
# split it in 1 second samples
len_new_sig = new_sig.shape[0]
num_samples = len_new_sig // sample_len
sample_wav_template = f"ltts_{wav_ID}_{{:03d}}.wav"
for i_sample in range(num_samples):
# cut the sample
sample_sig = new_sig[i_sample * sample_len : (i_sample + 1) * sample_len]
# get the name of the wav file
sample_name = sample_wav_template.format(i_sample)
# save the sample
sample_path = ltts_label_folder / sample_name
write_wav(sample_path, new_sr, sample_sig)
saved_samples += 1
# the ID of the sample
sample_id = f"{ltts_label}/{sample_name}"
# split in trainin/validation/testing
x = rng.random()
if x < test_fold_size:
word_test_list.append(sample_id)
elif x < test_fold_size + val_fold_size:
word_val_list.append(sample_id)
if saved_samples > max_num_samples:
break
# save the IDs
word_val_str = "\n".join(word_val_list)
val_list_path.write_text(word_val_str)
word_test_str = "\n".join(word_test_list)
test_list_path.write_text(word_test_str)
def save_wav(self, wav, path):
wav = wav * 32767 / max(0.01, np.max(np.abs(wav)))
write_wav(path, self.hparams.sample_rate, wav.astype(np.int16))
def split_ljspeech() -> None:
"""MAKEDOC: what is split_ljspeech doing?"""
logg = logging.getLogger(f"c.{__name__}.split_ljspeech")
# logg.setLevel("INFO")
logg.debug("Start split_ljspeech")
# the location of this file
this_file_folder = Path(__file__).parent.absolute()
logg.debug(f"this_file_folder: {this_file_folder}")
####################################################################################
# Find the list of sentences without training words inside #
####################################################################################
# the location of the original dataset
ljs_base_folder = Path("~").expanduser() / "audiodatasets" / "LJSpeech-1.1"
logg.debug(f"ljs_base_folder: {ljs_base_folder}")
ljs_wav_folder = ljs_base_folder / "wavs"
logg.debug(f"ljs_wav_folder: {ljs_wav_folder}")
# column names in the metadata file
column_names = ["wav_ID", "tra", "norm_tra"]
# load the metadata file
metadata_path = ljs_base_folder / "metadata.csv"
meta_df = pandas.read_csv(metadata_path,
sep="|",
header=0,
names=column_names,
index_col=False)
# shuffle it
meta_df = meta_df.sample(frac=1)
# remove NaN
meta_df = meta_df.dropna()
# show some rows to be sure
logg.debug(f"meta_df.head():\n{meta_df.head()}")
# get the words
train_words = words_types["all"]
# train_words = words_types["num"]
# count good sentences
good_sent = 0
skip_sent = 0
# find good sentences
good_IDs: ty.List[str] = []
for index_row, row in meta_df.iterrows():
norm_tra = row["norm_tra"]
# skip this row if one of the training words is in the sentence
found_train_word = False
for word in train_words:
if word in norm_tra:
# logg.debug(f"Found '{word}' in {norm_tra}")
found_train_word = True
skip_sent += 1
break
if found_train_word:
continue
good_sent += 1
# save the file ID
good_IDs.append(row["wav_ID"])
logg.debug(f"good_sent: {good_sent}")
logg.debug(f"skip_sent: {skip_sent}")
####################################################################################
# Create the samples #
####################################################################################
# where to save the results
# ljs_proc_folder = this_file_folder / "data_ljs_raw"
ljs_proc_folder = this_file_folder / "data_raw"
ljs_label = "_other_ljs"
ljs_label_folder = ljs_proc_folder / ljs_label
logg.debug(f"ljs_label_folder: {ljs_label_folder}")
if not ljs_label_folder.exists():
ljs_label_folder.mkdir(parents=True, exist_ok=True)
# the file to list the file names for the testing fold
val_list_path = ljs_proc_folder / "validation_list_ljspeech.txt"
word_val_list: ty.List[str] = []
test_list_path = ljs_proc_folder / "testing_list_ljspeech.txt"
word_test_list: ty.List[str] = []
# the sizes of the test and validation folds
val_fold_size = 0.1
test_fold_size = 0.1
# the target sr
new_sr = 16000
# how many samples to generate
max_num_samples = 5000
saved_samples = 0
# a random number generator to use
rng = np.random.default_rng(12345)
for wav_ID in good_IDs[:]:
wav_name = f"{wav_ID}.wav"
# load the original signal
orig_wav_path = ljs_wav_folder / wav_name
orig_sig, orig_sr = librosa.load(orig_wav_path, sr=None)
# logg.debug(f"orig_sig.shape: {orig_sig.shape} orig_sr: {orig_sr}")
# resample it to 16000 Hz
new_sig = librosa.resample(orig_sig, orig_sr, new_sr)
# logg.debug(f"new_sig.shape: {new_sig.shape}")
# split it in 1 second samples
len_new_sig = new_sig.shape[0]
num_samples = len_new_sig // new_sr
sample_wav_template = f"ljs_{wav_ID}_{{:03d}}.wav"
for i_sample in range(num_samples):
# cut the sample
sample_sig = new_sig[i_sample * new_sr:(i_sample + 1) * new_sr]
# get the name of the wav file
sample_name = sample_wav_template.format(i_sample)
# save the sample
sample_path = ljs_label_folder / sample_name
write_wav(sample_path, new_sr, sample_sig)
saved_samples += 1
# the ID of the sample
sample_id = f"{ljs_label}/{sample_name}"
# split in trainin/validation/testing
x = rng.random()
if x < test_fold_size:
word_test_list.append(sample_id)
elif x < test_fold_size + val_fold_size:
word_val_list.append(sample_id)
if saved_samples > max_num_samples:
break
# save the IDs
word_val_str = "\n".join(word_val_list)
val_list_path.write_text(word_val_str)
word_test_str = "\n".join(word_test_list)
test_list_path.write_text(word_test_str)
velocities=NSYNTH_VELOCITIES,
transpose=float(args.transpose)
)
synth.preload_notes(instrument=instrument['name'], source_type=instrument['source_type'])
instrument_folder = instrument['name']+'_'+instrument['source_type']
init_directory(os.path.join(args.audios_path, instrument_folder))
for mid in midifiles:
_, seq_name = os.path.split(mid)
output_name = os.path.join(args.audios_path, instrument_folder, os.path.splitext(seq_name)[0]+'.wav')
print("Instrumento: \t", instrument_folder)
print("Secuencia: \t", mid)
print("Salida: \t", output_name, '\n')
if(not os.path.isfile(output_name)):
audio, _ = synth.render_sequence(
sequence=str(mid),
instrument=instrument['name'],
source_type=instrument['source_type'],
preset=instrument['preset'],
playback_speed=float(args.playback_speed),
duration_scale=float(args.duration_rate),
)
if(DEFAULT_SAMPLING_RATE != NSYNTH_SAMPLE_RATE):
audio = librosa.core.resample(audio, NSYNTH_SAMPLE_RATE, DEFAULT_SAMPLING_RATE)
# write_audio(output_name, audio, DEFAULT_SAMPLING_RATE)
write_wav(output_name, DEFAULT_SAMPLING_RATE, np.array(32000.*audio, np.short))
def apiquery():
"""api interface"""
msg = str(json.loads(request.data).get('text', ''))
wav = tts.generate(msg)
write_wav('tmp.wav', hp.sr, wav)
return send_file('tmp.wav', attachment_filename='tmp.wav')
def save_wav(self, out_path, data):
write_wav(out_path, self.rate, data)
def preprocess_fsdd() -> None:
"""MAKEDOC: what is preprocess_fsdd doing?
Get the dataset with:
git clone https://github.com/Jakobovski/free-spoken-digit-dataset.git ~/free_spoken_digit_dataset
"""
logg = logging.getLogger(f"c.{__name__}.preprocess_fsdd")
# logg.setLevel("INFO")
logg.debug("Start preprocess_fsdd")
# the location of the original dataset
fsdd_base_folder = Path("~").expanduser().absolute()
fsdd_rec_folder = fsdd_base_folder / "free_spoken_digit_dataset" / "recordings"
logg.debug(f"fsdd_rec_folder: {fsdd_rec_folder}")
# the location of this file
this_file_folder = Path(__file__).parent.absolute()
logg.debug(f"this_file_folder: {this_file_folder}")
# where to save the results
# fsdd_proc_folder = this_file_folder / "data_fsdd_raw"
fsdd_proc_folder = this_file_folder / "data_raw"
logg.debug(f"fsdd_proc_folder: {fsdd_proc_folder}")
if not fsdd_proc_folder.exists():
fsdd_proc_folder.mkdir(parents=True, exist_ok=True)
# the file to list the file names for the testing fold
test_list_path = fsdd_proc_folder / "testing_list_fsdd.txt"
word_test_list = []
num2name = {
0: "fsdd_zero",
1: "fsdd_one",
2: "fsdd_two",
3: "fsdd_three",
4: "fsdd_four",
5: "fsdd_five",
6: "fsdd_six",
7: "fsdd_seven",
8: "fsdd_eight",
9: "fsdd_nine",
}
for num, name in num2name.items():
name_folder = fsdd_proc_folder / name
if not name_folder.exists():
name_folder.mkdir(parents=True, exist_ok=True)
# a random number generator to use
rng = np.random.default_rng(12345)
# the target sr
new_sr = 16000
for wav_path in tqdm(fsdd_rec_folder.iterdir()):
# logg.debug(f"wav_path: {wav_path}")
wav_name = wav_path.name
# logg.debug(f"wav_name: {wav_name}")
orig_sig, orig_sr = librosa.load(wav_path, sr=None)
# logg.debug(f"orig_sig.shape: {orig_sig.shape} orig_sr: {orig_sr}")
new_sig = librosa.resample(orig_sig, orig_sr, new_sr)
# logg.debug(f"new_sig.shape: {new_sig.shape}")
padded_sig = pad_signal(new_sig, new_sr)
# logg.debug(f"padded_sig.shape: {padded_sig.shape}")
# if there is space to move the signal around, roll it a bit
max_roll = (new_sr - new_sig.shape[0] - 50) // 2
if max_roll > 50:
sample_shift = rng.integers(-max_roll, max_roll)
rolled_sig = np.roll(padded_sig, sample_shift)
else:
rolled_sig = padded_sig
# the label of this word
word_label = num2name[int(wav_name[0])]
# save the processed file
name_folder = fsdd_proc_folder / word_label
new_path = name_folder / wav_name
write_wav(new_path, new_sr, rolled_sig)
# create the id for this word
word_id = f"{word_label}/{wav_name}"
word_test_list.append(word_id)
# save the IDs
word_test_str = "\n".join(word_test_list)
test_list_path.write_text(word_test_str)
ap.add_argument('--preload', required=False, default=True, help="Load all notes in memory before rendering for better performance (at least 1 GB of RAM is required)")
args = vars(ap.parse_args())
assert os.path.isdir(args['db']), 'Dataset not found in ' + args['db']
assert os.path.isfile(args['seq']), 'File ' + args['seq'] + ' not found.'
synth = NoteSynthesizer(
args['db'],
sr=NSYNTH_SAMPLE_RATE,
velocities=NSYNTH_VELOCITIES,
preload=args['preload']
)
if(args['preload']):
synth.preload_notes(args['instrument'], args['source_type'], int(args['preset']))
y, _ = synth.render_sequence(
sequence=args['seq'],
instrument=args['instrument'],
source_type=args['source_type'],
preset=int(args['preset']),
transpose=int(args['transpose']),
playback_speed=float(args['playback_speed']),
duration_scale=float(args['duration_scale'])
)
if(int(args['sr']) != NSYNTH_SAMPLE_RATE):
y = librosa.core.resample(y, NSYNTH_SAMPLE_RATE, int(args['sr']))
print("Saving audio output to", args['output'])
write_wav(args['output'], int(args['sr']), np.array(32000.*y, np.short))
x, y = generate_sine_wave(2)
plt.plot(x, y)
plt.show()
# Mezclando las señales de audio, superponiendolas
_, nice_tone = generate_sine_wave(400, SAMPLE_RATE, DURATION)
_, noise_tone = generate_sine_wave(4000, SAMPLE_RATE, DURATION)
noise_tone = noise_tone * 0.3 # atenuar 30%
mixed_tone = nice_tone + noise_tone
normalized_tone = np.int16((mixed_tone / mixed_tone.max()) * 32767)
plt.plot(normalized_tone[:1000])
plt.show()
# Guardando las muestras generadas en un archivo de audio
wav_path = os.path.join(BASE_DIR, 'signals_mixed.wav')
write_wav(wav_path, SAMPLE_RATE, normalized_tone)
# Usando FFT
# Number of samples in normalized_tone
N = SAMPLE_RATE * DURATION
yf = fft(normalized_tone)
xf = fftfreq(N, 1 / SAMPLE_RATE)
plt.plot(xf, np.abs(yf))
plt.xlabel('Frecuencia')
plt.title('Resultado FFT, reflejo en y=0')
print('FFT con reflejo en y=0')
plt.show()
# Volviendo la FFT mas rapida calculado solo la mitad positiva
# Note the extra 'r' at the front
yf = rfft(normalized_tone)
x_limited = x[:muestras_grafica]
y_limited = y[:muestras_grafica]
plt.plot(x_limited, y_limited)
plt.xlabel('Tiempo')
plt.ylabel('Amplitud')
plt.title('Señal pura con frecuencia f')
plt.show()
# Tono Deseado
tone = y
# Generando ruido
_, noise1 = generate_sine_wave(NOISE, FS, DURATION)
mixed_tone = tone + noise1
normalized = np.int16((mixed_tone / mixed_tone.max()) * 32767)
print('Generando audio con ruido')
write_wav(audio_filename, FS, normalized)
plt.plot(normalized[:muestras_grafica])
plt.xlabel('Tiempo')
plt.ylabel('Amplitud')
plt.title('Señal Tono + Ruido')
plt.show()
# FFT
# Note the extra 'r' at the front
N = FS * DURATION
yf = rfft(normalized)
xf = rfftfreq(N, 1 / FS)
plt.plot(xf, np.abs(yf))
plt.xlabel('Frecuencia')
plt.title('Cuadrante positivo')
plt.show()
