def recognise(self, mode='sphinx', marg=0.2):
# use the audio file as the audio source
r = srec.Recognizer()
if mode=='sphinx':
recogniser = r.recognize_sphinx
if mode=='google':
recogniser = r.recognise_google
for ii, (ts,te,lab) in enumerate(zip(self.tst,self.tend,self.label)):
tstart = st-marg
tend = end+marg
wo=w[int(tstart*fs):int(tend*fs)]
wavwrite('speech_sample.wav',fs,wo)
with srec.AudioFile('speech_sample.wav') as source:
audio = r.record(source) # read the entire audio file
try:
# for testing purposes, we're just using the default API key
# to use another API key, use `r.recognize_google(audio, key="GOOGLE_SPEECH_RECOGNITION_API_KEY")`
# instead of `r.recognize_google(audio)`
utt = recogniser(audio)
#utt = r.recognize_sphinx(audio)
self.label[ii] = utt
except srec.UnknownValueError:
print("Speech Recognition could not understand audio")
except srec.RequestError as e:
print("Could not request results {}".format(e))
def recognise(self, mode='sphinx', marg=0.2):
import speech_recognition as srec
# use the audio file as the audio source
r = srec.Recognizer()
if mode=='sphinx':
recogniser = r.recognize_sphinx
sys.stderr.write('Doing speech recognition with sphinx\n')
if mode=='google':
sys.stderr.write('Doing speech recognition with google\n')
recogniser = r.recognise_google
for ii, (ts,te,lab) in enumerate(zip(self.tst,self.tend,self.label)):
tstart = ts-marg
tend = te+marg
wo=self.x[int(tstart*self.sr):int(tend*self.sr)]
wavwrite('speech_sample.wav',self.sr,wo.astype('int16'))
with srec.AudioFile('speech_sample.wav') as source:
audio = r.record(source) # read the entire audio file
try:
# for testing purposes, we're just using the default API key
# to use another API key, use `r.recognize_google(audio, key="GOOGLE_SPEECH_RECOGNITION_API_KEY")`
# instead of `r.recognize_google(audio)`
utt = recogniser(audio)
#utt = r.recognize_sphinx(audio)
self.label[ii] = utt
sys.stderr.write('{}\n'.format(utt))
except srec.UnknownValueError:
sys.stderr.write("Speech Recognition could not understand audio\n")
except srec.RequestError as e:
sys.stderr.write("Could not request results {}\n".format(e))
def main():
try:
out_dir, in_dir = sys.argv[1], sys.argv[2]
except:
in_dir = '../../data/sc09'
out_dir = '../../data/sc09_wav'
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
tfrecord_fps = glob.glob(os.path.join(in_dir, '*.tfrecord'))
dataset = tf.data.TFRecordDataset(tfrecord_fps)
dataset = dataset.map(_mapper)
dataset = dataset.apply(tf.contrib.data.batch_and_drop_remainder(1))
x, y = dataset.make_one_shot_iterator().get_next()
x, y = x[0], y[0]
with tf.Session() as sess:
i = 0
while True:
try:
_x, _y = sess.run([x, y])
except:
break
_x *= 32767.
_x = np.clip(_x, -32767., 32767.)
_x = _x.astype(np.int16)
wavwrite(os.path.join(out_dir, '{}_{}.wav'.format(_y, str(i).zfill(5))), 16000, _x)
i += 1
def main(argv):
if str(argv[0]) == "help":
print("python genwave.py FREQ AMP DURATION FS NUMCH BITS f/i OUTPATH")
return
freq = float(argv[0]) if len(argv) > 0 else 440
amp = float(argv[1]) if len(argv) > 1 else 0.7
duration = float(argv[2]) if len(argv) > 2 else 30
fs = int(argv[3]) if len(argv) > 3 else 44100
nch = int(argv[4]) if len(argv) > 4 else 2
bits = int(argv[5]) if len(argv) > 5 else 16
f_or_i = argv[6] if len(argv) > 6 else "i"
outpath = str(argv[7]) if len(argv) > 7 else "out.wav"
if bits == 16:
sig = np.zeros([int(fs * duration), nch], dtype=np.int16)
elif bits == 32:
if f_or_i == "i":
sig = np.zeros([int(fs * duration), nch], dtype=np.int32)
else:
sig = np.zeros([int(fs * duration), nch], dtype=np.float32)
else:
print("invalid bit-width: 16/32 required")
return
sinewave = amp * np.sin(
np.arange(sig.shape[0]) * freq * 1.0 / fs * 2 * np.pi)
for ich in range(sig.shape[1]):
if f_or_i == "i":
sig[:, ich] = np.round(sinewave * (2**(bits - 1) - 1))
else:
sig[:, ich] = sinewave[:]
wavwrite(outpath, fs, sig)
def recover_samples_from_spectrum(logspectrum_stft, spectrum_phase, save_to):
abs_spectrum = np.exp(logspectrum_stft)
spectrum = abs_spectrum * (np.exp(1j * spectrum_phase))
istft_graph = tf.Graph()
with istft_graph.as_default():
num_fea = int(FLAGS.Fs * 0.025) / 2 + 1
frame_length = int(FLAGS.Fs * 0.025)
frame_step = int(FLAGS.Fs * 0.010)
stft_ph = tf.placeholder(tf.complex64, shape=(None, num_fea))
samples = tf.signal.inverse_stft(
stft_ph,
frame_length,
frame_step,
frame_length,
window_fn=tf.signal.inverse_stft_window_fn(
frame_step,
forward_window_fn=functools.partial(tf.signal.hann_window,
periodic=True)))
istft_sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True))
samples_ = istft_sess.run(samples, feed_dict={stft_ph: spectrum})
wavwrite(save_to, FLAGS.Fs, samples_)
return samples_
def getFilteredDataList(inputFileName, load = True, save = True):
frequencyFilterWidthMap = getNs()
result = []
[originalSampleRate, original] = wavread(inputFileName)
for cutOffFrequency in frequencyFilterWidthMap:
outputFileName = inputFileName + '_noisedfiltered_' + str(cutOffFrequency) + '.wav'
if os.path.isfile(outputFileName) and load:
print("Loading file ", outputFileName, " from disk." )
[sampleRate, x] = wavread(outputFileName)
if sampleRate != originalSampleRate:
raise ValueError("Sample rate of file ", outputFileName, " does not eaqual the sample rate of",
" the original file " , inputFileName)
else:
windowSize = frequencyFilterWidthMap[cutOffFrequency]
print('Generating noisedfiltered ', cutOffFrequency, ' data, with windowSize ', windowSize)
x = running_mean(original, windowSize).astype('int16')
x = x + np.random.normal(0, 10, len(x)).astype('int16') # to add noise.
if save:
wavwrite(outputFileName, originalSampleRate, x)
print("saved: ", outputFileName)
if len(x) != len(original):
raise ValueError("Filtering the wav file failed. Original input is ", len(original), " long",
"but the filtered data is " , len(x) , " long.")
result.append(x)
return (result ,originalSampleRate)
def test_signal_normalize(tmp_path):
sig = np.sin(2*np.pi*np.linspace(0,1,50))
sig1 = sig * .8
sig2 = sig * .4
rms_out1 = 0.700 # Should be very close to full amplitude
rms_out2 = 0.350 # Should be very close to half amplitude
with tempfile.TemporaryDirectory() as tmpdirin:
infile1 = os.path.join(tmpdirin, 'file1.wav')
infile2 = os.path.join(tmpdirin, 'file2.wav')
wavwrite(infile1, 50, sig1)
wavwrite(infile2, 50, sig2)
with tempfile.TemporaryDirectory() as tmpdirout:
psylab.signal.normalize(tmpdirin, ext='wav', outdir=tmpdirout, relative=False)
fs,out1 = wavread(os.path.join(tmpdirout, 'file1.wav'))
fs,out2 = wavread(os.path.join(tmpdirout, 'file2.wav'))
#print("sig: {:}".format(np.sqrt(np.mean(np.square(sig)))))
#print("sig1: {:}".format(np.sqrt(np.mean(np.square(sig1)))))
#print("sig2: {:}".format(np.sqrt(np.mean(np.square(sig2)))))
#print("out1: {:}".format(np.sqrt(np.mean(np.square(out1)))))
#print("out2: {:}".format(np.sqrt(np.mean(np.square(out2)))))
np.testing.assert_allclose(rms_out1, np.sqrt(np.mean(np.square(out1))), rtol=1e-3)
np.testing.assert_allclose(rms_out2, np.sqrt(np.mean(np.square(out2))), rtol=1e-3)
def gen(path, endian="little"):
endian_tag = "<" if endian == "little" else ">"
print("parse the file \"{}\"".format(path))
tags = path.split("_")
tags = map(lambda s: s.replace(".pcm", ""), tags)
fs_tag = filter(lambda s: s.startswith("sr"), tags)
if len(fs_tag) > 0:
fs = int(fs_tag[0][2:])
else:
print("filename \"{}\" does not match the format.".format(path))
return
ch_tag = filter(lambda s: s.startswith("ch"), tags)
if len(ch_tag) > 0:
num_ch = int(ch_tag[0][2:])
else:
print("filename \"{}\" does not match the format.".format(path))
return
format_tag = filter(lambda s: s.startswith("format"), tags)
if len(format_tag) > 0:
proc_format = int(format_tag[0][6:])
else:
print("filename \"{}\" does not match the format.".format(path))
return
if not proc_format in [1, 5, 6]:
print("the format must be pcm16/pcm24/pcmfloat.")
return
with open(path, "rb") as f:
raw_data = f.read()
if proc_format == 1:
signal = np.array(struct.unpack(
"{}{}h".format(endian_tag,
len(raw_data) / 2), raw_data),
dtype=np.int16)
elif proc_format == 5:
signal = np.array(struct.unpack(
"{}{}f".format(endian_tag,
len(raw_data) / 4), raw_data),
dtype=np.float32)
elif proc_format == 6:
data_len = len(raw_data) / 3
data = ""
for x in range(data_len):
data += raw_data[x * 3 + 1:x * 3 + 3]
signal = np.array(struct.unpack(
"{}{}h".format(endian_tag,
len(data) / 2), data),
dtype=np.int16)
if num_ch > 1:
signal = np.reshape(signal, (len(signal) / num_ch, num_ch))
output_path = "{}.wav".format(path)
wavwrite(filename=output_path, rate=fs, data=signal)
print("generate the file \"{}\"".format(output_path))
def main():
sound_filename = sys.argv[1]
frequency = float(sys.argv[2])
duration = float(sys.argv[3])
amplitude = float(sys.argv[4])
sampling_rate = int(sys.argv[5])
data = sine(frequency, duration, amplitude, sampling_rate)
wavwrite(sound_filename, sampling_rate, data)
def _wav_write(wav_fp, fs, wav_f, normalize=False):
if normalize:
wav_f_max = wav_f.max()
if wav_f_max != 0.0:
wav_f /= wav_f.max()
wav_f = np.clip(wav_f, -1.0, 1.0)
wav = (wav_f * 32767.0).astype(np.int16)
wavwrite(wav_fp, fs, wav)
def transcribe(self, input_audio, model_path=None, output="./"):
"""Transcribe frame-level fundamental frequency of vocal from the given audio.
Parameters
----------
input_audio: Path
Path to the wav audio file.
model_path: Path
Path to the trained model or the transcription mode. If given a path, should be
the folder that contains `arch.yaml`, `weights.h5`, and `configuration.yaml`.
output: Path (optional)
Path for writing out the extracted vocal f0. Default to current path.
Returns
-------
f0: txt
The transcribed f0 of the vocal contour in Hz.
See Also
--------
omnizart.cli.vocal_contour.transcribe: The coressponding command line entry.
"""
if not os.path.isfile(input_audio):
raise FileNotFoundError(
f"The given audio path does not exist. Path: {input_audio}")
logger.info("Loading model...")
model, model_settings = self._load_model(model_path)
logger.info("Extracting feature...")
feature = extract_cfp_feature(
input_audio,
hop=model_settings.feature.hop_size,
win_size=model_settings.feature.window_size,
down_fs=model_settings.feature.sampling_rate)
logger.info("Predicting...")
f0 = inference(feature[:, :, 0],
model,
timestep=model_settings.training.timesteps)
agg_f0 = aggregate_f0_info(f0, t_unit=model_settings.feature.hop_size)
timestamp = np.arange(len(f0)) * model_settings.feature.hop_size
wav = sonify.pitch_contour(timestamp,
f0,
model_settings.feature.sampling_rate,
amplitudes=0.5 * np.ones(len(f0)))
output = self._output_midi(output, input_audio, verbose=False)
if output is not None:
write_agg_f0_results(agg_f0, f"{output}_f0.csv")
wavwrite(f"{output}_trans.wav",
model_settings.feature.sampling_rate, wav)
logger.info("Text and Wav files have been written to %s",
os.path.abspath(os.path.dirname(output)))
logger.info("Transcription finished")
return agg_f0
def main(**kwargs):
outfile = kwargs['outfile'][0]
infile = kwargs['infile']
# print "Filtering %s to %s" % (infile, outfile)
data, rate = ffmpeg_load_audio(infile, 44100, True, dtype=np.float32)
wavwrite('test.wav', 44100, data)
filtered_data = butter_bandpass_filter(data, 100.0, 3000.0, 44100)
wavwrite(outfile, 44100, filtered_data)
def process_filtering(self,
sig_float_filtered,
write=False,
output_file_name=None):
self.filtered = True
self.sig_int = float2pcm(sig_float_filtered)
self.sig_float = sig_float_filtered
if write:
wavwrite(output_file_name, file.sr, file.sig_int)
def generate_audio(self, fake_r, fake_i, order):
audio_dir = './audio/' + str(order) + '/'
if not os.path.isdir(audio_dir):
os.makedirs(audio_dir)
for i in range(self.batch_size):
audio_fp = os.path.join(audio_dir, '{}.wav'.format(str(i)))
fake = fake_r[i, 0] + 1j * fake_i[i, 0]
audio = spec_to_wav(fake)
wavwrite(audio_fp, self.fs, audio)
print("Done generating audio :)")
def resample(path_in, path_out=None, sampling_rate=16000):
# load and resampling
y, sr = librosa.load(path_in, sr=sampling_rate)
# save add '-16k' as suffix if path_out is None
if path_out is None:
root, ext = os.path.splitext(path_in)
path_out= root + '-16k' + ext
wavwrite( path_out, sr , ( y * 2 ** 15).astype(np.int16))
print ('save wav file', path_out)
def get_wav(self, *, resample=True):
'''Retrieves io.BytesIO() packed with `.wav` contents'''
result = self.resample_fs(self.BULLSHITWAVNUMBER) if resample \
else self.copy()
data = result.normdata(dtype=np.int16)
bytes_io = io.BytesIO()
wavwrite(bytes_io, result._fs, data)
return bytes_io
def tx1_to_wav(tx1_fp, out_fp, midi_downsample_rate=None):
if midi_downsample_rate == 0:
midi_downsample_rate = None
print('(Rate {}) {}->{}'.format(midi_downsample_rate, tx1_fp, out_fp))
with open(tx1_fp, 'r') as f:
tx1 = f.read()
midi = tx1_to_midi(tx1)
wav = nesmdb.convert.midi_to_wav(midi, midi_downsample_rate)
wavwrite(out_fp, 44100, wav)
print('Done: {}'.format(wav.shape))
return True
def main(argv):
del argv
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
config = tf.contrib.tpu.RunConfig(
cluster=tpu_cluster_resolver,
model_dir=FLAGS.model_dir,
tpu_config=tf.contrib.tpu.TPUConfig(
num_shards=FLAGS.num_shards,
iterations_per_loop=FLAGS.iterations_per_loop))
# Set module-level global variable so that model_fn and input_fn can be
# identical for each different kind of dataset and model
global dataset, model
dataset = bias_input
model = bias_model
# TPU-based estimator used for TRAIN and EVAL
est = tf.contrib.tpu.TPUEstimator(model_fn=model_fn,
use_tpu=FLAGS.use_tpu,
config=config,
train_batch_size=FLAGS.batch_size,
eval_batch_size=FLAGS.batch_size)
# CPU-based estimator used for PREDICT (generating images)
cpu_est = tf.contrib.tpu.TPUEstimator(model_fn=model_fn,
use_tpu=False,
config=config,
predict_batch_size=_NUM_VIZ_AUDIO)
current_step = estimator._load_global_step_from_checkpoint_dir(
FLAGS.model_dir) # pylint: disable=protected-access,line-too-long
tf.logging.info('Starting training for %d steps, current step: %d' %
(FLAGS.train_steps, current_step))
# Render some generated images
G_z = cpu_est.predict(input_fn=noise_input_fn)
G_z = [p['generated_audio'][:, :] for p in G_z]
G_z = np.array(G_z)
preview_dir = './preview'
if not os.path.isdir(preview_dir):
os.makedirs(preview_dir)
for i in range(len(G_z)):
audio = np.int16(G_z[i] / np.max(np.abs(G_z[i])) * 32767)
preview_fp = os.path.join(
preview_dir, '{}_{}_{}.wav'.format(str(i % 10), str(current_step),
str(i)))
wavwrite(preview_fp, _FS, audio)
tf.logging.info('Finished generating images')
def recover_samples_from_spectrum(logspectrum_stft, spectrum_phase, save_to=None):
abs_spectrum = np.exp(logspectrum_stft)
spectrum = abs_spectrum * (np.exp(1j * spectrum_phase))
istft_graph = tf.Graph()
with istft_graph.as_default():
stft_ph = tf.placeholder(tf.complex64, shape=(None, 201))
samples = tf.contrib.signal.inverse_stft(stft_ph, 400, 160, 400)
istft_sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
samples_ = istft_sess.run(samples, feed_dict={stft_ph: spectrum})
if save_to:
wavwrite(save_to, 16000, samples_)
return samples_
def save_wav(path0, data, sr=44100):
#
print('file ', path0)
amplitude = np.iinfo(np.int16).max
max_data = np.amax(np.abs(data)) # normalize, max level is 16bit full bit
if max_data < (1.0 / amplitude):
max_data = 1.0
try:
wavwrite(path0, sr,
np.array((amplitude / max_data) * data, dtype=np.int16))
except:
print('error: wavwrite ', path0)
sys.exit()
def record_audio(self, out_path, time_allowance):
self.ids.record_button.disabled = True
self.ids.reading_speed_slider.disabled = True
self.ids.goto_text_input.disabled = True
recording = self.mic.record(samplerate=self.fs,
numframes=int(time_allowance * self.fs),
channels=1)
recording = recording / numpy.max(numpy.abs(recording))
wavwrite(out_path, self.fs, recording) # Save as WAV file
self.recording_indicator = ""
self.ids.record_button.disabled = False
self.ids.reading_speed_slider.disabled = False
self.ids.goto_text_input.disabled = False
self.load_next_sentence()
def test_signal_equate(tmp_path):
sig1 = np.sin(2*np.pi*np.linspace(0,1,50))
sig2 = sig1 * .8
with tempfile.TemporaryDirectory() as tmpdirin:
infile1 = os.path.join(tmpdirin, 'file1.wav')
infile2 = os.path.join(tmpdirin, 'file2.wav')
wavwrite(infile1, 50, sig1)
wavwrite(infile2, 50, sig2)
with tempfile.TemporaryDirectory() as tmpdirout:
psylab.signal.equate(tmpdirin, ext='wav', outdir=tmpdirout, relative=False)
fs,out1 = wavread(os.path.join(tmpdirout, 'file1.wav'))
fs,out2 = wavread(os.path.join(tmpdirout, 'file2.wav'))
np.testing.assert_allclose(sig2, out1)
np.testing.assert_allclose(sig2, out2)
np.testing.assert_allclose(np.sqrt(np.mean(np.square(sig1)))*.8, np.sqrt(np.mean(np.square(out2))))
def filter(input_path, output_path, lowcut=100.0, highcut=3000.0, rate=44100):
"""
Filters a wav file to get rid of frequencies not present in human speech. Band pass filter.
:param input_path: the input wav path
:param output_path: the output wav path
:param lowcut: the lowest frequency to accept
:param highcut: the highest frequency to accept
:param rate: the sampling frequency
:type rate: Number
:param force: call recursively if the input_path and match_path lengths vary greatly (not in betwee 0.5 and 2.0)
:returns: -1 if a file does not exist
"""
if check_file_paths([input_path]) == -1:
return -1
data, rate = ffmpeg_load_audio(input_path, 44100, True, dtype=np.float32)
filtered_data = butter_bandpass_filter(data, lowcut, highcut, rate)
wavwrite(output_path, 44100, filtered_data)
def epoch_save(self, samples, dir_name, epoch):
"""
Makes a mag/phase figure using _get_image and plotimage from
samples and saves it as a .png file. Samples are also transformed
to audio signals and saved as a .wav file. All files are saved
into a subdirectory with the epoch number as its name.
Parameters
----------
samples : ndarray
Array of mag/phase tensors you want to save. Needs to hold at
least 64 tensors.
dir_name : string
Name of the target directory.
epoch : int
Specifies the name of directory containing the saved figure
and .wav ffiles . It is advisable that it is max 3 digits
long, since the name format is "%00%d".
"""
# Create new epoch directory
ep_dir_name = dir_name + '/%s' % str(epoch).zfill(3) + '/'
if not os.path.exists(ep_dir_name):
os.makedirs(ep_dir_name)
# Create a 8x8 grid of mag/phase images
batch_size, _, width, _ = samples.shape
samples_magphase = np.zeros((batch_size, width, width))
# samples_magphase[:, :width // 2, :] = np.squeeze(samples[:, :, :, 0])
samples_magphase[:, :width // 2, :] = samples[:, :, :,
0] # Spectrogram
samples_magphase[:, width // 2:, :] = samples[:, :, :, 1] # Phase
# Save the image
fig = self.plotimage(samples_magphase)
plt.savefig(ep_dir_name + 'mag_phase.png', bbox_inches='tight')
plt.close(fig)
# Save sound samples
for i, magphase in enumerate(samples):
audio = self._image_to_audio(magphase)
wavwrite(ep_dir_name + "%s.wav" % str(i).zfill(3), int(self.rate),
audio / self.rate)
def handle_signals(mixedpath, noisepospath, noisenegpath):
try:
# Read Wavs
mixedsamples = read_wav(mixedpath)
noisepossamples = read_wav(noisepospath)
noisenegsamples = read_wav(noisenegpath)
# Normalize
max_scale = max(abs(mixedsamples)+0.000001)
mixedsamples = mixedsamples / max_scale
wavwrite('/home/user/Desktop/N_HANS_Github/N_HANS___Selective_Noise/audio_examples/mixed_normalised', 16000, mixedsamples)
noisepossamples = noisepossamples / max_scale
noisenegsamples = noisenegsamples / max_scale
mixedsamples = mixedsamples.astype(np.float32)
noisepossamples = noisepossamples.astype(np.float32)
noisenegsamples = noisenegsamples.astype(np.float32)
# Cut the end to have an exact number of frames
if (len(mixedsamples) - 400) % 160 != 0:
mixedsamples = mixedsamples[:-((len(mixedsamples) - 400) % 160)]
nsepos = noisepossamples
nseneg = noisenegsamples
while len(mixedsamples) - len(nsepos) > 0: # Make noise longer
diff = len(mixedsamples) - len(nsepos)
nsepos = np.concatenate([nsepos, noisepossamples[:diff]], axis=0)
while len(mixedsamples) - len(nseneg) > 0: # Make noise longer
diff = len(mixedsamples) - len(nseneg)
nseneg = np.concatenate([nseneg, noisenegsamples[:diff]], axis=0)
if len(mixedsamples) - len(noisepossamples) < 0: # Make noise shorter
nsepos = noisepossamples[:len(mixedsamples)]
if len(mixedsamples) - len(noisenegsamples) < 0: # Make noise shorter
nseneg = noisenegsamples[:len(mixedsamples)]
noisepossamples = nsepos
noisenegsamples = nseneg
return noisepossamples, noisenegsamples, mixedsamples
except:
print('error in threads')
print(mixedpath, noisepospath, noisenegpath)
def filter(input_path, output_path, lowcut=100.0, highcut=3000.0, rate=44100):
"""
Filters a wav file to get rid of frequencies not present in human speech. Band pass filter.
:param input_path: the input wav path
:param output_path: the output wav path
:param lowcut: the lowest frequency to accept
:param highcut: the highest frequency to accept
:param rate: the sampling frequency
:type rate: Number
:param force: call recursively if the input_path and match_path lengths vary greatly (not in betwee 0.5 and 2.0)
:returns: -1 if a file does not exist
"""
if check_file_paths([input_path]) == -1:
return -1
data, rate = ffmpeg_load_audio(input_path, 44100, True, dtype=np.float32)
filtered_data = butter_bandpass_filter(data, lowcut, highcut, rate)
# return filtered_data
wavwrite(output_path, 44100, filtered_data)
def runTest():
dir = 'ressources/'
files = [
dir + 'test1.mpg', dir + 'test2.mpg', dir + 'test3.mpg',
dir + 'test4.mpg', dir + 'test5.mpg'
]
sg = SoundGenerator(sound)
j = 0
fileList = []
for i in files:
objs = k.kanadeHarris(i, sound)
data, fs = sg.soundGenerationForVideoPurpose(objs)
wavwrite("soundTest" + str(j) + ".wav", fs, data)
genNewVideo('ressources/test' + str(j + 1) + '.mpg_out.avi',
"soundTest" + str(j) + ".wav", j)
fileList.append('res/output' + str(j) + '.avi')
j += 1
concatRes(fileList)
def reconstructFromSTFT(spectrum_phase, logspectrum_stft, save_to):
abs_spectrum = np.exp(logspectrum_stft)
spectrum_phase = np.array(spectrum_phase)
spectrum = abs_spectrum * (np.exp(1j * spectrum_phase))
istft_graph = tf.Graph()
with istft_graph.as_default():
num_fea = FFT_LENGTH // 2 + 1
frame_length = WIN_SAMPLES
frame_step = HOP_SAMPLES
stft_ph = tf.placeholder(tf.complex64, shape=(None, num_fea))
samples = tf.signal.inverse_stft(stft_ph, frame_length, frame_step,
FFT_LENGTH)
istft_sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True))
samples_ = istft_sess.run(samples, feed_dict={stft_ph: spectrum})
wavwrite(save_to, SAMPLING_RATE, samples_)
return samples_
def average(*args):
"""
Averages multiple wav files together. Accomplishes this by performing fast fourier transforms on the data, averaging those arrays, and then performing an inverse fast fourier transform.
:param args: array of wav paths with the output being the first path and the rest being inputs
:returns: -1 if it fails or if it cannot find the paths
:Example:
>>> import speechprocessing
>>> speechprocessing.average('output.wav', 'input_one.wav', 'input_two.wav')
"""
if len(args) < 2:
print 'Invalid number of arguments'
return -1
output_path = args[0]
input_paths = args[1:]
processed_wav_data = []
if check_file_paths(input_paths) == -1:
return -1
for path in input_paths:
data, rate = ffmpeg_load_audio(path, 44100, True, dtype=np.float32)
filtered_data = butter_bandpass_filter(data, 100.0, 3000.0, 44100)
processed_wav_data.append(filtered_data)
fft_data = []
for data in processed_wav_data:
fft_data.append(np.fft.rfft(data))
# Adding a * before an array of arrays makes it zip array-wise
# .. or something. Nobody really knows how or why this works
zipped_data = zip(*fft_data)
mean_data = map(np.mean, zipped_data)
# Reverse real fft
averaged = np.fft.irfft(mean_data)
wavwrite(output_path, 44100.0, averaged)
def average(*args):
"""
Averages multiple wav files together. Accomplishes this by performing fast fourier transforms on the data, averaging those arrays, and then performing an inverse fast fourier transform.
:param args: array of wav paths with the output being the first path and the rest being inputs
:returns: -1 if it fails or if it cannot find the paths
:Example:
>>> import speechprocessing
>>> speechprocessing.average('output.wav', 'input_one.wav', 'input_two.wav')
"""
if len(args) < 2:
print 'Invalid number of arguments'
return -1
output_path = args[0]
input_paths = args[1:]
processed_wav_data = []
if check_file_paths(input_paths) == -1:
return -1
for path in input_paths:
data, rate = ffmpeg_load_audio(path, 44100, True, dtype=np.float32)
filtered_data = butter_bandpass_filter(data, 100.0, 3000.0, 44100)
processed_wav_data.append(filtered_data)
fft_data = []
for data in processed_wav_data:
fft_data.append(np.fft.rfft(data))
# Adding a * before an array of arrays makes it zip array-wise
# .. or something. Nobody really knows how or why this works
zipped_data = zip(*fft_data)
mean_data = map(np.mean, zipped_data)
# Reverse real fft
averaged = np.fft.irfft(mean_data)
wavwrite(output_path, 44100.0, averaged)
def log_batch(self,
pr_batch,
t_batch,
mix_batch,
mixture_rec=None):
"""!
:param pr_batch: Reconstructed wavs: Torch Tensor of size:
batch_size x num_sources x length_of_wavs
:param t_batch: Target wavs: Torch Tensor of size:
batch_size x num_sources x length_of_wavs
:param mix_batch: Batch of the mixtures: Torch Tensor of size:
batch_size x 1 x length_of_wavs
:param mixture_rec: Batch of the reconstructed mixtures: Torch Tensor of
size: batch_size x 1 x length_of_wavs
"""
mixture = mix_batch.detach().cpu().numpy()
true_sources = t_batch.detach().cpu().numpy()
pred_sources = pr_batch.detach().cpu().numpy()
for b_ind in range(self.bs):
self.log_example(pred_sources[b_ind],
true_sources[b_ind],
mixture[b_ind],
title='bind_{}'.format(b_ind))
if mixture_rec is not None:
mixture_rec_np = mixture_rec.detach().cpu().numpy()
for b_ind in range(self.bs):
rec_mix_name = "bind_{}_rec_mix.wav".format(b_ind)
rec_mix_wav = (mixture_rec_np[b_ind][0] /
(np.max(np.abs(mixture_rec_np[b_ind][0])) +
10e-8))
wavwrite(os.path.join(self.dirpath, rec_mix_name),
self.fs,
rec_mix_wav)
def main(**kwargs):
outfile = kwargs['outfile'][0]
infile = kwargs['infile']
print "Filtering %s to %s" % (infile, outfile)
rate, sound_samples = wavread(infile)
mono = True
if 'ndarray' in str(type(sound_samples[0])):
mono = False
# data,r = ffmpeg_load_audio('32but.wav', 44100, True, dtype=np.float32)
rate, sound_samples = ffmpeg_load_audio(infile, rate, mono, dtype=np.float32)
fs = 44100.0
lowcut = 100.0
highcut = 3000.0
# b,a = butter_bandpass(lowcut, highcut, fs, 5)
# filtered = lfilter(b, a, sound_samples)
# filtered = butter_bandpass_filter(sound_samples, lowcut, highcut, fs, 5)
# filtered = butter_bandpass_filter_two(sound_samples, lowcut, highcut, fs, 5)
wavwrite(outfile, rate, sound_samples)
def main():
sig = 3
offset = 0
rate, data, nrate, ndata = get_all_data()
timelength = data.shape[0] / rate
times = np.linspace(0., timelength, data.shape[0])
kernel = signal.gaussian(data.size, sig)
f, pxx, nf, npxx, convolved, nconvolved = get_power_spectra(
kernel, sig, PLOT_FLAG)
# Test smoothing
gaussian_smoothing_test(PLOT_FLAG)
# Apply custom low-pass filter
filtered = low_pass(convolved, nconvolved, f, times, PLOT_FLAG)
# Recover the final signal
recovered_signal = recover(filtered, data, times, offset, PLOT_FLAG)
# Save this signal into a wav file
recovered_signal.real.astype(np.int16)
wavwrite("reconstructed.wav", rate, recovered_signal)
def log_example(self,
pred_sources,
true_sources,
mixture,
title=''):
mix_name = "{}_mix.wav".format(title)
mix_wav = (mixture[0] /
(np.max(np.abs(mixture[0])) + 10e-8))
wavwrite(os.path.join(self.dirpath, mix_name),
self.fs,
mix_wav)
for s_ind in range(self.n_sources):
true_s_name = "{}_true_s{}.wav".format(title, s_ind)
rec_s_name = "{}_rec_s{}.wav".format(title, s_ind)
rec_wav = (pred_sources[s_ind] /
(np.max(np.abs(pred_sources[s_ind])) + 10e-8))
true_wav = (true_sources[s_ind] /
(np.max(np.abs(true_sources[s_ind])) + 10e-8))
wavwrite(os.path.join(self.dirpath, true_s_name),
self.fs,
true_wav)
wavwrite(os.path.join(self.dirpath, rec_s_name),
self.fs,
rec_wav)
recons_vt_psd = lpc.lpc2psd(lpc_coef, g, fft_size)
recons_psd *= recons_vt_psd
recons_psds.append(recons_psd)
recons_vt_psds.append(recons_vt_psd)
recons_psds = np.array(recons_psds)
recons_vt_psds = np.array(recons_vt_psds)
df = pysptk.synthesis.AllZeroDF(ORDER)
synthesizer = pysptk.synthesis.Synthesizer(df, int(fs * 0.005))
x_glottal_res_zerodf = synthesizer.synthesis(x, lpcs / gains)
df = pysptk.synthesis.AllZeroDF(3)
synthesizer = pysptk.synthesis.Synthesizer(df, int(fs * 0.005))
x_res_zerodf = synthesizer.synthesis(x_glottal_res_zerodf,
glottal_lpcs / glottal_gains)
wavwrite('x_glottal_res_zerodf.wav', fs,
(x_glottal_res_zerodf * 2**15).astype(np.int16))
wavwrite('x_res_zerodf.wav', fs, (x_res_zerodf * 2**15).astype(np.int16))
y = synthesisRequiem.get_waveform(x_res_zerodf,
np.transpose(recons_psds, [1, 0]),
dat['temporal_positions'], dat['f0'],
dat['fs'])
y_from_glottal = synthesisRequiem.get_waveform(
x_glottal_res_zerodf, np.transpose(recons_vt_psds, [1, 0]),
dat['temporal_positions'], dat['f0'], dat['fs'])
wavwrite('x_recons_zerodf.wav', fs, (y * 2**15).astype(np.int16))
wavwrite('x_recons_glottal_zerodf.wav', fs,
(y_from_glottal * 2**15).astype(np.int16))
output_filepath = sys.argv[1]
input_filepaths = sys.argv[2:]
fs = 44100.0
lowcut = 100.0 # Low pass cutoff
highcut = 3000.0 # High pass cutoff
processed_wav_data = []
for path in input_filepaths:
data, rate = ffmpeg_load_audio(path, 44100, True, dtype=np.float32)
filtered_data = butter_bandpass_filter(data, 100.0, 3000.0, 44100)
processed_wav_data.append(filtered_data)
fft_data = []
for data in processed_wav_data:
fft_data.append(np.fft.rfft(data))
# Adding a * before an array of arrays makes it zip array-wise
# .. or something. Nobody really knows how or why this works
zipped_data = zip(*fft_data)
mean_data = map(np.mean, zipped_data)
# Reverse real fft
f = np.fft.irfft(mean_data)
wavwrite(output_filepath, fs, f)
def save_vgmwav(wav_fp, wav): wav *= 32767. wav = np.clip(wav, -32768., 32767.) wav = wav.astype(np.int16) wavwrite(wav_fp, 44100, wav)
isIce=allsicRS>thresh
isIce[nanMask]=np.nan
probIce=np.nanmean(isIce, axis=-1)
p25=[]
p75=[]
for yit in range(len(boxPos)):
p25.append(np.percentile(nSIF[yit,:], 25))
p75.append(np.percentile(nSIF[yit,:], 75))
p25_1850= np.percentile(nSIF1850, 25)
p75_1850= np.percentile(nSIF1850, 75)
data1=(sic[:,0]>15)*1.
data2=(sic1850>15)*1.
wavwrite('/Volumes/Pitcairn/seaicePPF/northernHemisphere/figures/OWnoise1850-2100.'+key+'.wav', 365*8, data1)
wavwrite('/Volumes/Pitcairn/seaicePPF/northernHemisphere/figures/OWnoise1850BG'+key+'.wav', 365*8, data2)
fig8, ax8 = plt.subplots(1, 2, sharey=True, sharex=False, num=None, figsize=(7,3.5), dpi=300, facecolor='w', edgecolor='w')
fig8.sca(ax8[0])
plt.tick_params(axis='both', which='major', labelsize=6)
plt.tick_params(axis='both', which='minor', labelsize=6)
fig8.sca(ax8[1])
plt.tick_params(axis='both', which='major', labelsize=6)
plt.tick_params(axis='both', which='minor', labelsize=6)
fig8.patch.set_alpha(0.0)
#ax5.set_aspect('equal', 'datalim')
fig8.sca(ax8[1])
plt.imshow(meanAllsic.T, vmin=0, vmax=100, cmap=cmap)
#plt.colorbar()
# sound_time = f.nframes*1.0/f.samplerate
# sound_data = f.read_frames(f.nframes)
# samples_to_take = int(math.floor(sound_time * display_sample_rate))
# time_step_for_samples = f.samplerate*1.0/display_sample_rate
# wave = []
# for i in xrange(samples_to_take):
# frame_offset = i * time_step_for_samples
# if num_channels == 1:
# wave.append(sound_data[frame_offset])
# else:
# wave.append(sound_data[frame_offset][0])
rate, wave = wavread(infile)
wavwrite('test.wav', rate, wave)
(freq, amp) = get_component_frequencies(wave)
# print type(s)
# with open('data.txt', 'a') as textOutputFile:
# for line in amp:
# textOutputFile.write(str(line))
# textOutputFile.write(',')
# Only plot first 4000 Hz
hz=4000
freq = freq[0:hz]
amp = amp[0:hz]
fig = pylab.figure()
def sound(var):
from scipy.io.wavfile import write as wavwrite
scaled = np.int16(var/np.max(np.abs(var)) * 32767)
stmp=asarray(scaled,dtype=np.int16)
wavwrite('stmp.wav',8820,stmp)
sysfileopen("stmp.wav")
import numpy as np
import wave
from scipy.io.wavfile import write as wavwrite
import struct
LENWAV = 20000 # Must be <= 40000 / SAMPLEFACTOR for one-second wav files
SAMPLEFACTOR = 4
img = np.zeros((1, 1, LENWAV, 1), dtype=int)
waveFile = wave.open("../data/pitbull.wav", "rb")
for i in range(0, LENWAV * SAMPLEFACTOR):
waveData = waveFile.readframes(1)
if i % SAMPLEFACTOR == 0:
sound = struct.unpack("<h", waveData)
img[0, 0, i / SAMPLEFACTOR, 0] = sound[0]
print "NEW SAMPLE RATE", waveFile.getframerate() / SAMPLEFACTOR
wavwrite(
"../output/test.wav",
waveFile.getframerate() / SAMPLEFACTOR,
img.flatten() / float(np.max(np.abs(img.flatten()), axis=0)),
)
def deemphasis(signal):
return lfilter([1, 0.70], 1, signal)
def rms(signal):
return sqrt(mean(power(signal, 2)))
if __name__ == "__main__":
fs, data = wavread('Mann.wav')
data = array(data, dtype=double)
data /= amax(absolute(data))
data = decimate(data, 4)
fs = round(fs/4)
block_len = 0.032
overlap = 0.5
order = 16
out = vocode(data, fs, block_len, overlap, order)
wavwrite('vocoded.wav', fs, array(out/amax(absolute(out)) * (2**15-1), dtype=int16))
figure()
plot(data)
figure()
plot(out)
show()
# ideas:
# use reduce(ola, map(process, array))
# http://stackoverflow.com/questions/6657820/python-convert-an-iterable-to-a-stream
def process_filtering(self, sig_float_filtered, write=False, output_file_name = None):
self.filtered = True
self.sig_int = float2pcm(sig_float_filtered)
self.sig_float = sig_float_filtered
if write:
wavwrite(output_file_name, file.sr, file.sig_int)
def main():
import argparse
import os
import sys
import traceback
from tqdm import tqdm
parser = argparse.ArgumentParser()
conversion_to_types = {
# VGM simplifiers
'vgm_simplify': ('.vgm', '.simp.vgm'),
'vgm_shorten': ('.vgm', '.short.vgm'),
# NES disassembly raw
'vgm_to_ndr': ('.vgm', '.ndr.pkl'),
'ndr_to_txt': ('.ndr.pkl', '.ndr.txt'),
'txt_to_ndr': ('.ndr.txt', '.ndr.pkl'),
'ndr_to_vgm': ('.ndr.pkl', '.ndr.vgm'),
# NES disassembly functional
'vgm_to_ndf': ('.vgm', '.ndf.pkl'),
'ndf_to_txt': ('.ndf.pkl', '.ndf.txt'),
'txt_to_ndf': ('.ndf.txt', '.ndf.pkl'),
'ndf_to_vgm': ('.ndf.pkl', '.ndf.vgm'),
# NES language modeling format
'vgm_to_nlm': ('.vgm', '.nlm.pkl'),
'nlm_to_vgm': ('.nlm.pkl', '.nlm.vgm'),
# NES-MDB score formats
'ndf_to_exprsco': ('.ndf.pkl', '.exprsco.pkl'),
'ndf_to_midi': ('.ndf.pkl', '.mid'),
'exprsco_to_seprsco': ('.exprsco.pkl', '.seprsco.pkl'),
'exprsco_to_blndsco': ('.exprsco.pkl', '.blndsco.pkl'),
# WAV converters
'vgm_to_wav': ('.vgm', '.wav'),
'ndr_to_wav': ('.ndr.pkl', '.wav'),
'ndf_to_wav': ('.ndf.pkl', '.wav'),
'nlm_to_wav': ('.nlm.pkl', '.wav'),
'midi_to_wav': ('.mid', '.mid.wav'),
'exprsco_to_wav': ('.exprsco.pkl', '.exprsco.wav'),
'seprsco_to_wav': ('.seprsco.pkl', '.seprsco.wav'),
'blndsco_to_wav': ('.blndsco.pkl', '.blndsco.wav'),
}
conversion_to_kwargs = {
'vgm_simplify': ['vgm_simplify_nop1', 'vgm_simplify_nop2', 'vgm_simplify_notr', 'vgm_simplify_nono'],
'vgm_shorten': ['vgm_shorten_start', 'vgm_shorten_nmax'],
'ndf_to_exprsco': ['ndf_to_exprsco_rate'],
'midi_to_wav': ['midi_to_wav_rate'],
}
parser.add_argument('conversion', type=str, choices=conversion_to_types.keys())
parser.add_argument('fps', type=str, nargs='+')
parser.add_argument('--out_dir', type=str)
parser.add_argument('--skip_verify', action='store_true', dest='skip_verify')
parser.add_argument('--vgm_shorten_start', type=int)
parser.add_argument('--vgm_shorten_nmax', type=int)
parser.add_argument('--vgm_simplify_nop1', action='store_true', dest='vgm_simplify_nop1')
parser.add_argument('--vgm_simplify_nop2', action='store_true', dest='vgm_simplify_nop2')
parser.add_argument('--vgm_simplify_notr', action='store_true', dest='vgm_simplify_notr')
parser.add_argument('--vgm_simplify_nono', action='store_true', dest='vgm_simplify_nono')
parser.add_argument('--ndf_to_exprsco_rate', type=float)
parser.add_argument('--midi_to_wav_rate', type=float)
parser.set_defaults(
conversion=None,
fps=None,
out_dir=None,
skip_verify=False,
vgm_shorten_start=None,
vgm_shorten_nmax=1024,
vgm_simplify_nop1=False,
vgm_simplify_nop2=False,
vgm_simplify_notr=False,
vgm_simplify_nono=False,
ndf_to_exprsco_rate=None,
midi_to_wav_rate=None)
args = parser.parse_args()
in_type, out_type = conversion_to_types[args.conversion]
fps = args.fps
if len(fps) > 1 and args.out_dir is None:
raise Exception('Must specify output directory for batch mode')
if len(fps) == 1 and args.out_dir is None:
out_fps = [fps[0].replace(in_type, out_type)]
else:
out_fns = [os.path.basename(fp).replace(in_type, out_type) for fp in fps]
out_fps = [os.path.join(args.out_dir, fn) for fn in out_fns]
if os.path.exists(args.out_dir):
print 'WARNING: Output directory {} already exists'.format(args.out_dir)
else:
os.makedirs(args.out_dir)
for in_fp, out_fp in tqdm(zip(fps, out_fps)):
if not args.skip_verify:
_verify_type(in_fp, in_type)
_verify_type(out_fp, out_type)
# Load input file
in_ext = in_type.split('.')[-1]
if in_ext == 'pkl':
with open(in_fp, 'rb') as f:
in_file = pickle.load(f)
elif in_ext in ['mid', 'vgm']:
with open(in_fp, 'rb') as f:
in_file = f.read()
elif in_ext == 'txt':
with open(in_fp, 'r') as f:
in_file = f.read()
else:
raise NotImplementedError('Input extension .{} not recognized'.format(in_ext))
kwargs = {}
if args.conversion in conversion_to_kwargs:
kwargs = {kw:getattr(args, kw) for kw in conversion_to_kwargs[args.conversion]}
try:
out_file = globals()[args.conversion](in_file, **kwargs)
except:
print '-' * 80
print in_fp
traceback.print_exc()
continue
# Save output file
out_ext = out_type.split('.')[-1]
if out_ext == 'pkl':
with open(out_fp, 'wb') as f:
pickle.dump(out_file, f)
elif out_ext in ['mid', 'vgm']:
with open(out_fp, 'wb') as f:
f.write(out_file)
elif out_ext == 'txt':
with open(out_fp, 'w') as f:
f.write(out_file)
elif out_ext == 'wav':
wav = out_file
wav *= 32767.
wav = np.clip(wav, -32767., 32767.)
wav = wav.astype(np.int16)
wavwrite(out_fp, 44100, wav)
else:
raise NotImplementedError('Output extension .{} not recognized'.format(out_ext))
sigs = mixsounds_ica(g,dg)
sigs = np.array(sigs,dtype = 'uint8')
plot(sigs[0],len(sigs[0]))
plot(sigs[1],len(sigs[1]))
plot(sigs[2],len(sigs[2]))
plot(sigs[3],len(sigs[3]))
plot(sigs[4],len(sigs[4]))
plot(sigs[5],len(sigs[5]))
plot(sigs[6],len(sigs[6]))
plot(sigs[7],len(sigs[7]))
plot(sigs[8],len(sigs[8]))
# write in a file
if 'output' not in os.listdir('.'):
os.mkdir('output')
for i in range(sigs.shape[0]):
wavwrite('output/unmixedsound'+`i`+'.wav',8000,sigs[i])
# used to test the quality of the functions (part 1)
elif sys.argv[1] == 'convergence':
mean = 0
N = 100
N_n = N
for i in range(N):
s1,s2,n = square_cos_ica(g,dg)
if n >= 0:
mean += n
#if algo diverged then don't count it
else:
N_n = N_n - 1
print(mean)
print("The mean of the number of steps untill convergence is " +`mean/N_n`)