InferenceInterfaces/ControllableInterface.py

import os

import torch

from Architectures.ControllabilityGAN.GAN import GanWrapper
from InferenceInterfaces.ToucanTTSInterface import ToucanTTSInterface
from Utility.storage_config import MODELS_DIR


class ControllableInterface:

 def __init__(self, gpu_id="cpu", available_artificial_voices=1000):
 if gpu_id == "cpu":
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
 else:
 os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
 os.environ["CUDA_VISIBLE_DEVICES"] = f"{gpu_id}"
 self.device = "cuda" if gpu_id != "cpu" else "cpu"
 self.model = ToucanTTSInterface(device=self.device, tts_model_path="Meta", language="eng")
 self.wgan = GanWrapper(os.path.join(MODELS_DIR, "Embedding", "embedding_gan.pt"), device=self.device)
 self.generated_speaker_embeds = list()
 self.available_artificial_voices = available_artificial_voices

 def read(self,
 prompt,
 voice_seed,
 prosody_creativity,
 duration_scaling_factor,
 pause_duration_scaling_factor,
 pitch_variance_scale,
 energy_variance_scale,
 emb_slider_1,
 emb_slider_2,
 emb_slider_3,
 emb_slider_4,
 emb_slider_5,
 emb_slider_6,
 loudness_in_db
 ):
 self.wgan.set_latent(voice_seed)
 controllability_vector = torch.tensor([emb_slider_1,
 emb_slider_2,
 emb_slider_3,
 emb_slider_4,
 emb_slider_5,
 emb_slider_6], dtype=torch.float32)
 embedding = self.wgan.modify_embed(controllability_vector)
 self.model.set_utterance_embedding(embedding=embedding)

 phones = self.model.text2phone.get_phone_string(prompt)
 if len(phones) > 1800:
 prompt = "Your input was too long. Please try either a shorter text or split it into several parts."

 print(prompt)
 wav, sr, fig = self.model(prompt,
 input_is_phones=False,
 duration_scaling_factor=duration_scaling_factor,
 pitch_variance_scale=pitch_variance_scale,
 energy_variance_scale=energy_variance_scale,
 pause_duration_scaling_factor=pause_duration_scaling_factor,
 return_plot_as_filepath=True,
 prosody_creativity=prosody_creativity,
 loudness_in_db=loudness_in_db)
 return sr, wav, fig