def _get_wav_and_melspec(wav_file, length, is_training=True):
'''
the range of values of wav is [-1, 1].
'''
wav = read_wav(wav_file, sr=hp.signal.sr)
wav = trim_wav(wav)
# divide wav into chunks that have the given length and one is randomly selected in training, but first chunk in generation.
n_clips = math.ceil(len(wav) / length) if is_training else 1
idx = random.randrange(n_clips)
start, end = length * idx, length * (idx + 1)
wav = wav[start:end]
assert (len(wav) <= length)
wav = fix_length(wav, length) # padding in case of last chunk.
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
wav = np.expand_dims(wav, -1)
return wav, melspec.astype(np.float32)
def do_inference(num_tests, concurrency=1):
channel = implementations.insecure_channel(host, int(port))
stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
coord = _Coordinator(num_tests, concurrency)
for _ in range(num_tests):
# dummy audio
duration, sr, n_fft, win_length, hop_length, n_mels, max_db, min_db = 4, 16000, 512, 512, 128, 80, 35, -55
filename = librosa.util.example_audio_file()
wav = read_wav(filename, sr=sr, duration=duration)
mel = wav2melspec_db(wav, sr, n_fft, win_length, hop_length, n_mels)
mel = normalize_db(mel, max_db=max_db, min_db=min_db)
mel = mel.astype(np.float32)
mel = np.expand_dims(mel, axis=0) # single batch
n_timesteps = sr / hop_length * duration + 1
# build request
request = predict_pb2.PredictRequest()
request.model_spec.name = 'voice_vector'
request.model_spec.signature_name = 'predict'
request.inputs['x'].CopyFrom(
tf.contrib.util.make_tensor_proto(mel,
shape=[1, n_timesteps, n_mels]))
coord.throttle()
# send asynchronous response (recommended. use this.)
result_future = stub.Predict.future(request, 10.0) # timeout
result_future.add_done_callback(_create_rpc_callback(coord))
# send synchronous response (NOT recommended)
# result = stub.Predict(request, 5.0)
coord.wait_all_done()
def get_random_wav_and_label(self, tar_wavfiles, ntar_wavfiles):
"""
:return: wav: raw wave. float32. shape=(t, ),
label: 1 if target, 0 otherwise. int32.
melspec: mel-spectrogram. float32. shape=(t, n_mels)
"""
wavfiles, label = (
tar_wavfiles,
self.tar_labels) if np.random.sample(1) <= self.tar_ratio else (
ntar_wavfiles, self.ntar_labels)
wavfile = wavfiles[np.random.randint(0, len(wavfiles))]
if type(wavfile) == bytes:
wavfile = wavfile.decode()
if wavfile.endswith('arr'): # pyarrow format
wav = read_wav_from_arr(wavfile)
else:
wav = read_wav(wavfile, sr=hp.signal.sr)
wav = trim_wav(wav)
wav = crop_random_wav(wav, self.length)
wav = augment_volume(wav)
wav = fix_length(wav, self.length) # padding
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
melspec = np.float32(melspec)
label = np.float32(label)
return wav, melspec, label
def get_data(self):
while True:
speaker_id = random.choice(list(self.speaker_dict.keys()))
wav = self._load_random_wav(speaker_id)
mel_spec = wav2melspec_db(wav, hp.signal.sr, hp.signal.n_fft,
hp.signal.win_length,
hp.signal.hop_length, hp.signal.n_mels)
mel_spec = normalize_db(mel_spec,
max_db=hp.signal.max_db,
min_db=hp.signal.min_db)
yield wav, mel_spec, speaker_id
def get_random_wav(self, wavfile):
"""
:param: wavfile: a raw wave file.
:return: wav: raw wave. float32. shape=(t, ),
melspec: mel-spectrogram. float32. shape=(t, n_mels),
wavfile: the raw wave file.
"""
wav = read_wav(wavfile, sr=hp.signal.sr)
wav = trim_wav(wav)
wav = fix_length(wav, self.length) # crop from the beginning.
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
melspec = np.float32(melspec)
return wav, melspec, wavfile
def _get_wav_and_melspec(wav_file, length=None, is_training=True):
wav = read_wav(wav_file, sr=hp.signal.sr)
wav = trim_wav(wav)
if length:
n_clips = math.ceil(len(wav) / length) if is_training else 1
idx = random.randrange(n_clips)
start, end = length * idx, length * (idx + 1)
wav = wav[start:end]
assert (len(wav) <= length)
wav = fix_length(wav, length) # padding
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
wav = np.expand_dims(wav, -1)
return wav, melspec.astype(np.float32)
