def convert_worker(
config: Config,
wrapper: VoiceChangerStreamWrapper,
acoustic_converter: AcousticConverter,
super_resolution: SuperResolution,
audio_config: AudioConfig,
queue_input: Queue,
queue_output: Queue,
):
wrapper.voice_changer_stream.voice_changer = VoiceChanger(
super_resolution=super_resolution,
acoustic_converter=acoustic_converter,
)
start_time = 0
time_length = audio_config.convert_chunk / audio_config.rate
while True:
in_feature: AcousticFeatureWrapper = queue_input.get()
wrapper.voice_changer_stream.add_in_feature(
start_time=start_time,
feature_wrapper=in_feature,
frame_period=audio_config.frame_period,
)
start_time += time_length
out_feature = wrapper.convert_next(time_length=time_length)
queue_output.put(out_feature)
def voice_changer(self):
if self._voice_changer is None:
acoustic_converter, super_resolution = self.models
self._voice_changer = VoiceChanger(
acoustic_converter=acoustic_converter,
super_resolution=super_resolution,
output_sampling_rate=self.out_sampling_rate,
)
return self._voice_changer
def convert_worker(
acoustic_converter: AcousticConverter,
super_resolution: SuperResolution,
time_length: float,
extra_time: float,
input_silent_threshold: float,
queue_input: Queue,
queue_output: Queue,
acquired_lock: Lock,
):
logger = logging.getLogger('convert')
init_logger(logger)
logging.info('convert worker')
chainer.global_config.enable_backprop = False
chainer.global_config.train = False
stream = ConvertStream(voice_changer=VoiceChanger(
super_resolution=super_resolution,
acoustic_converter=acoustic_converter,
threshold=input_silent_threshold,
))
stream_wrapper = StreamWrapper(stream=stream, extra_time=extra_time)
acquired_lock.release()
start_time = extra_time
while True:
item: Item = queue_input.get()
start = time.time()
in_feature: AcousticFeatureWrapper = item.item
stream.add(
start_time=start_time,
data=in_feature,
)
start_time += time_length
out_feature = stream_wrapper.process_next(time_length=time_length)
item.item = out_feature
queue_output.put(item)
logger.debug(f'{item.index}: {time.time() - start}')
def check(
input_path: Path,
input_time_length: int,
output_path: Path,
input_statistics_path: Path,
target_statistics_path: Path,
stage1_model_path: Path,
stage1_config_path: Path,
stage2_model_path: Path,
stage2_config_path: Path,
):
ac_config = create_config(stage1_config_path)
sr_config = create_sr_config(stage2_config_path)
input_rate = ac_config.dataset.acoustic_param.sampling_rate
output_rate = sr_config.dataset.param.voice_param.sample_rate
realtime_vocoder = RealtimeVocoder(
acoustic_param=ac_config.dataset.acoustic_param,
out_sampling_rate=output_rate,
extract_f0_mode=VocodeMode.WORLD,
)
realtime_vocoder.create_synthesizer(
buffer_size=1024,
number_of_pointers=16,
)
f0_converter = F0Converter(
input_statistics=input_statistics_path,
target_statistics=target_statistics_path,
)
ac_config = ac_config
sr_config = sr_config
acoustic_converter = AcousticConverter(
ac_config,
stage1_model_path,
f0_converter=f0_converter,
out_sampling_rate=output_rate,
)
super_resolution = SuperResolution(
sr_config,
stage2_model_path,
)
voice_changer = VoiceChanger(
acoustic_converter=acoustic_converter,
super_resolution=super_resolution,
output_sampling_rate=output_rate,
)
encode_stream = EncodeStream(vocoder=realtime_vocoder)
convert_stream = ConvertStream(voice_changer=voice_changer)
decode_stream = DecodeStream(vocoder=realtime_vocoder)
num_data = input_time_length
time_length = 1
def _load_wave_and_split(time_length: float = 1):
length = round(time_length * input_rate)
wave, _ = librosa.load(str(input_path), sr=input_rate)
return [
wave[i * length:(i + 1) * length]
for i in range(len(wave) // length)
]
def _add(_stream: BaseStream, _datas):
for i, data in zip(range(num_data), _datas):
_stream.add(start_time=i * time_length, data=data)
def _split_process(_stream: BaseStream, _extra_time: float):
return [
_stream.process(start_time=i * time_length,
time_length=time_length,
extra_time=_extra_time) for i in range(num_data)
]
def _join_process(_stream: BaseStream, _extra_time: float):
return _stream.process(start_time=0,
time_length=time_length * num_data,
extra_time=_extra_time)
def _process_all_stream(
_streams: Tuple[BaseStream, BaseStream, BaseStream],
_datas,
_split_flags: Tuple[bool, bool, bool],
_extra_times: Tuple[float, float, float],
):
for stream, split_flag, extra_time in zip(_streams, _split_flags,
_extra_times):
_add(stream, _datas)
if split_flag:
_datas = _split_process(stream, _extra_time=extra_time)
else:
_datas = [_join_process(stream, _extra_time=extra_time)]
return _datas
def _concat_and_save(_waves, _path: Path):
wave = numpy.concatenate(_waves).astype(numpy.float32)
librosa.output.write_wav(str(_path), wave, output_rate)
def _remove(_streams: Tuple[BaseStream, BaseStream, BaseStream]):
for stream in _streams:
stream.remove(end_time=num_data)
waves = _load_wave_and_split(time_length=time_length)[:num_data]
encode_stream = encode_stream
convert_stream = convert_stream
decode_stream = decode_stream
streams = (encode_stream, convert_stream, decode_stream)
datas = _process_all_stream(streams,
waves,
_split_flags=(True, True, True),
_extra_times=(0, 1, 0))
_concat_and_save(datas, output_path)
_remove(streams)
rate=config.dataset.acoustic_param.sampling_rate,
chunk=config.dataset.acoustic_param.sampling_rate,
vocoder_buffer_size=config.dataset.acoustic_param.sampling_rate // 16,
out_norm=4.5,
)
frame_period = config.dataset.acoustic_param.frame_period
vocoder = RealtimeVocoder(
acoustic_param=config.dataset.acoustic_param,
out_sampling_rate=audio_config.rate,
buffer_size=audio_config.vocoder_buffer_size,
number_of_pointers=16,
)
voice_changer = VoiceChanger(
super_resolution=super_resolution,
acoustic_converter=acoustic_converter,
)
voice_changer_stream = VoiceChangerStream(
sampling_rate=audio_config.rate,
frame_period=acoustic_converter._param.frame_period,
order=acoustic_converter._param.order,
in_dtype=numpy.float32,
)
voice_changer_stream.voice_changer = voice_changer
voice_changer_stream.vocoder = vocoder
wrapper = VoiceChangerStreamWrapper(
voice_changer_stream=voice_changer_stream,
extra_time_pre=0.2,