tts_hf.py

import sounddevice as sd
import torch

if __name__ == "__main__":
    from base import BaseMouth
else:
    from .base import BaseMouth


class Mouth_hf(BaseMouth):
    def __init__(
        self,
        model_id="kakao-enterprise/vits-vctk",
        device="cpu",
        forward_params={"speaker_id": 10},
        player=sd,
        wait=True,
        logger=None,
    ):
        from transformers import pipeline

        self.pipe = pipeline("text-to-speech", model=model_id, device=device)
        self.device = device
        self.forward_params = forward_params
        super().__init__(
            sample_rate=self.pipe.sampling_rate, player=player, wait=wait, logger=logger
        )

    def run_tts(self, text):
        with torch.no_grad():
            # inputs = self.tokenizer(text, return_tensors="pt")
            # inputs = inputs.to(self.device)
            # output = self.model(**inputs, speaker_id=self.speaker_id).waveform[0].to('cpu')
            output = self.pipe(text, forward_params=self.forward_params)
            self.sample_rate = output["sampling_rate"]
            return output["audio"][0]


if __name__ == "__main__":
    device = "cuda" if torch.cuda.is_available() else "cpu"
    mouth = Mouth_hf(
        model_id="kakao-enterprise/vits-vctk",
        device=device,
        forward_params={"speaker_id": 10},
    )

    text = (
        "If there's one thing that makes me nervous about the future of self-driving cars, it's that they'll "
        "replace human drivers.\nI think there's a huge opportunity to make human-driven cars safer and more "
        "efficient. There's no reason why we can't combine the benefits of self-driving cars with the ease of use "
        "of human-driven cars."
    )
    print(text)
    mouth.say_multiple(text, lambda x: False)
    sd.wait()