Spaces:
Running
Running
File size: 1,461 Bytes
574ab7e |
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 |
import json
import torch
import numpy as np
from .demo_speaker_embeddings import DemoSpeakerEmbeddings
class DemoRandomAnonymizer:
def __init__(self, device, vec_type='xvector', in_scale=False):
self.device = device
self.vec_type = vec_type
self.in_scale = in_scale
self.dim_ranges = None
self.embedding_extractor = DemoSpeakerEmbeddings(vec_type=self.vec_type, device=self.device)
def load_parameters(self, model_dir):
with open(model_dir / 'settings.json') as f:
settings = json.load(f)
self.vec_type = settings['vec_type'] if 'vec_type' in settings else self.vec_type
self.in_scale = settings.get('in_scale', self.in_scale)
if self.in_scale:
with open(model_dir / 'stats_per_dim.json') as f:
dim_ranges = json.load(f)
self.dim_ranges = [(v['min'], v['max']) for k, v in sorted(dim_ranges.items(), key=lambda x: int(x[0]))]
def anonymize_embedding(self, audio, sr):
speaker_embedding = torch.tensor(self.embedding_extractor.extract_vector_from_audio(wave=audio, sr=sr))
if self.dim_ranges:
anon_vec = torch.tensor([np.random.uniform(*dim_range) for dim_range in self.dim_ranges]).to(self.device)
else:
mask = torch.zeros(speaker_embedding.shape[0]).float().random_(-40, 40).to(self.device)
anon_vec = speaker_embedding * mask
return anon_vec
|