def __call__(self, inputs: np.array) -> Tuple[np.array, int, List[str]]:
"""
Args:
inputs (:obj:`np.array`):
The raw waveform of audio received. By default sampled at `self.sampling_rate`.
The shape of this array is `T`, where `T` is the time axis
Return:
A :obj:`tuple` containing:
- :obj:`np.array`:
The return shape of the array must be `C'`x`T'`
- a :obj:`int`: the sampling rate as an int in Hz.
- a :obj:`List[str]`: the annotation for each out channel.
This can be the name of the instruments for audio source separation
or some annotation for speech enhancement. The length must be `C'`.
"""
_inputs = torch.from_numpy(inputs).unsqueeze(0)
sample = S2THubInterface.get_model_input(self.task, _inputs)
text = S2THubInterface.get_prediction(self.task, self.model,
self.generator, sample)
if self.tts_model is None:
return np.zeros((0, )), self.sampling_rate, [text]
else:
tts_sample = TTSHubInterface.get_model_input(self.tts_task, text)
wav, sr = TTSHubInterface.get_prediction(self.tts_task,
self.tts_model,
self.tts_generator,
tts_sample)
return wav.unsqueeze(0).numpy(), sr, [text]
def from_pretrained(
cls,
model_name_or_path,
checkpoint_file="model.pt",
data_name_or_path=".",
config_yaml="config.yaml",
**kwargs,
):
from fairseq import hub_utils
x = hub_utils.from_pretrained(
model_name_or_path,
checkpoint_file,
data_name_or_path,
archive_map=cls.hub_models(),
config_yaml=config_yaml,
**kwargs,
)
return S2THubInterface(x["args"], x["task"], x["models"][0])