Spaces:
Running
Running
File size: 4,031 Bytes
b3fa29f e8958d3 b3fa29f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 |
import os
import librosa.display as lbd
import matplotlib.pyplot as plt
import soundfile
import torch
from InferenceInterfaces.InferenceArchitectures.InferenceFastSpeech2 import FastSpeech2
from InferenceInterfaces.InferenceArchitectures.InferenceHiFiGAN import HiFiGANGenerator
from Preprocessing.ArticulatoryCombinedTextFrontend import ArticulatoryCombinedTextFrontend
from Preprocessing.ArticulatoryCombinedTextFrontend import get_language_id
from Preprocessing.ProsodicConditionExtractor import ProsodicConditionExtractor
class Meta_FastSpeech2(torch.nn.Module):
def __init__(self, device="cpu"):
super().__init__()
model_name = "Meta"
language = "en"
self.device = device
self.text2phone = ArticulatoryCombinedTextFrontend(language=language, add_silence_to_end=True)
checkpoint = torch.load(os.path.join("Models", f"FastSpeech2_{model_name}", "best.pt"), map_location='cpu')
self.phone2mel = FastSpeech2(weights=checkpoint["model"]).to(torch.device(device))
self.mel2wav = HiFiGANGenerator(path_to_weights=os.path.join("Models", "HiFiGAN_combined", "best.pt")).to(torch.device(device))
self.default_utterance_embedding = checkpoint["default_emb"].to(self.device)
self.phone2mel.eval()
self.mel2wav.eval()
self.lang_id = get_language_id(language)
self.to(torch.device(device))
def set_utterance_embedding(self, path_to_reference_audio):
wave, sr = soundfile.read(path_to_reference_audio)
self.default_utterance_embedding = ProsodicConditionExtractor(sr=sr).extract_condition_from_reference_wave(wave).to(self.device)
def set_language(self, lang_id):
"""
The id parameter actually refers to the shorthand. This has become ambiguous with the introduction of the actual language IDs
"""
self.text2phone = ArticulatoryCombinedTextFrontend(language=lang_id, add_silence_to_end=True, silent=True)
self.lang_id = get_language_id(lang_id).to(self.device)
def forward(self, text, view=False, durations=None, pitch=None, energy=None, phones = False):
with torch.inference_mode():
if phones is False:
phones = self.text2phone.string_to_tensor(text).to(torch.device(self.device))
else:
phones = self.text2phone.string_to_tensor(text, input_phonemes=True).to(torch.device(self.device))
mel, durations, pitch, energy = self.phone2mel(phones,
return_duration_pitch_energy=True,
utterance_embedding=self.default_utterance_embedding,
durations=durations,
pitch=pitch,
energy=energy)
mel = mel.transpose(0, 1)
wave = self.mel2wav(mel)
if view:
from Utility.utils import cumsum_durations
fig, ax = plt.subplots(nrows=2, ncols=1)
ax[0].plot(wave.cpu().numpy())
lbd.specshow(mel.cpu().numpy(),
ax=ax[1],
sr=16000,
cmap='GnBu',
y_axis='mel',
x_axis=None,
hop_length=256)
ax[0].yaxis.set_visible(False)
ax[1].yaxis.set_visible(False)
duration_splits, label_positions = cumsum_durations(durations.cpu().numpy())
ax[1].set_xticks(duration_splits, minor=True)
ax[1].xaxis.grid(True, which='minor')
ax[1].set_xticks(label_positions, minor=False)
ax[1].set_xticklabels(self.text2phone.get_phone_string(text))
ax[0].set_title(text)
plt.subplots_adjust(left=0.05, bottom=0.1, right=0.95, top=.9, wspace=0.0, hspace=0.0)
plt.show()
return wave
|