Spaces:
Running on Zero
Running on Zero
File size: 10,802 Bytes
cc348e2 | 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 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 | # Description: loads pretrained or self trained embedding model and
# extracts the embedding vectors from audio and calculates
# the average (np.mean) for each bundle (file_list)
# Output: *.pt file
# Author: Victor Garcia, Lorenz Gutscher
# year: 2022
from speechbrain.pretrained import EncoderClassifier
import torch
import os
from tqdm import tqdm
import numpy as np
from glob import glob
from numpy import trim_zeros
from sklearn.manifold import TSNE
import matplotlib.pyplot as plt
# import seaborn as sns
import soundfile as sf
from AudioPreprocessor import AudioPreprocessor
from pathlib import Path
device=torch.device("cpu")
# language_id = EncoderClassifier.from_hparams(source="speechbrain/lang-id-commonlanguage_ecapa",
# run_opts={"device": str(device)},
# savedir="pretrained_models/lang-id-commonlanguage_ecapa")
language_id = EncoderClassifier.from_hparams(source="/data/vokquant/speechbrain/recipes/CommonLanguage/lang_id/results/ECAPA-TDNN/1986/save/ckpt/",
run_opts={"device": str(device)},
hparams_file="/data/vokquant/speechbrain/recipes/CommonLanguage/lang_id/inference/hyperparams_inference.yaml",
savedir="pretrained_models/lang-id-commonlanguage_ecapa")
# It takes as input a list containing the paths to all wavs of a same language or variety and returns the average embedding (centroid (np.mean))
def get_language_embedding(filelist,path_to_wavs,language_id):
temp=[]
_, sr = sf.read(os.path.join(path_to_wavs,filelist[0]))
print(sr)
ap = AudioPreprocessor(input_sr=sr, output_sr=16000, melspec_buckets=80, hop_length=256, n_fft=1024, cut_silence=False, device=device)
for wav in tqdm(filelist):
wave, sr = sf.read(os.path.join(path_to_wavs,wav))
# audio = language_id.load_audio(os.path.join(path_to_wavs,wav),savedir='/data/vokquant/IMS-Toucan_lang_emb/Preprocessing/softlinks_embedding')
try:
norm_wave = ap.audio_to_wave_tensor(normalize=True, audio=wave)
except ValueError:
continue
norm_wave = torch.tensor(trim_zeros(norm_wave.numpy()))
embedding = language_id.encode_batch(norm_wave.cpu().detach())
# add embedding of each wav to temp:
temp.append(embedding)
stack=np.stack(temp)
# average all embeddings:
embedding=np.mean(stack,axis=0)
print(embedding)
return embedding
WASS_corpora_generation = False
MLS_generation_spanish = False
MLS_generation_english = False
MLS_generation_dutch = False
MLS_generation_polish = False
MLS_generation_italian = False
MLS_generation_portuguese = False
MLS_generation_french = True
MLS_generation_german = False
root_path = '../../data/mls/mls_'
# generate spanish(sp) language_embedding average:
if MLS_generation_spanish==True:
language_name='spanish'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1] # only 1 file per folder is used [0:1]
print(files)
# Once you have the input list to the function, you call it
embedding=get_language_embedding(files,path_to_wavs,language_id)
# Finally, you save the embedding in the computer
torch.save(embedding,language_name+'_emb_trained.pt')
# generate english(en) language_embedding average:
if MLS_generation_english==True:
language_name='english'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1]
print(files)
embedding=get_language_embedding(files,path_to_wavs,language_id)
torch.save(embedding,language_name+'_emb_trained.pt')
# generate dutch(du) language_embedding average:
if MLS_generation_dutch==True:
language_name='dutch'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1]
print(files)
embedding=get_language_embedding(files,path_to_wavs,language_id)
torch.save(embedding,language_name+'_emb_trained.pt')
# generate polish(pl) language_embedding average:
if MLS_generation_polish==True:
language_name='polish'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1]
print(files)
embedding=get_language_embedding(files,path_to_wavs,language_id)
torch.save(embedding,language_name+'_emb_trained.pt')
# generate italian(it) language_embedding average:
if MLS_generation_italian==True:
language_name='italian'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1]
print(files)
embedding=get_language_embedding(files,path_to_wavs,language_id)
torch.save(embedding,language_name+'_emb_trained.pt')
# generate portuguese(pt) language_embedding average:
if MLS_generation_portuguese==True:
language_name='portuguese'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1]
print(files)
embedding=get_language_embedding(files,path_to_wavs,language_id)
torch.save(embedding,language_name+'_emb_trained.pt')
# generate german(de) language_embedding average:
if MLS_generation_german==True:
language_name='german'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1]
print(files)
embedding=get_language_embedding(files,path_to_wavs,language_id)
torch.save(embedding,language_name+'_emb_trained.pt')
# generate french(fr) language_embedding average:
if MLS_generation_french==True:
language_name='french'
path_to_wavs = ''
files = []
for (dirpath, dirnames, filenames) in os.walk(root_path+language_name+'/train/'):
files += [os.path.join(dirpath, file) for file in filenames if file[-5:]=='.flac'][0:1]
print(files)
embedding=get_language_embedding(files,path_to_wavs,language_id)
torch.save(embedding,language_name+'_emb_trained.pt')
if WASS_corpora_generation==True:
path_to_wavs = '/data/vokquant/data/aridialect/aridialect_wav16000/'
#This is used to generate the input list to the function get_language_embedding()
########################################################################################
for paths,dirs,files in tqdm(os.walk(path_to_wavs)):
at_files = [x for x in files if x.split('_')[1] == 'at']
vd_files = [x for x in files if x.split('_')[1] == 'vd']
goi_files = [x for x in files if x.split('_')[1] == 'goi']
ivg_files = [x for x in files if x.split('_')[1] == 'ivg']
########################################################################################
#This is another possibility
########################################################################################
# at_files = [str(path) for path in Path(path_to_wavs).resolve().glob("**/*_at_*.wav")]
# vd_files = [str(path) for path in Path(path_to_wavs).resolve().glob("**/*_vd_*.wav")]
# goi_files = [str(path) for path in Path(path_to_wavs).resolve().glob("**/*_goi_*.wav")]
# ivg_files = [str(path) for path in Path(path_to_wavs).resolve().glob("**/*_ivg_*.wav")]
########################################################################################
at_embedding=get_language_embedding(filelist=at_files,path_to_wavs=path_to_wavs,language_id=language_id)
vd_embedding=get_language_embedding(vd_files,path_to_wavs,language_id)
goi_embedding=get_language_embedding(goi_files,path_to_wavs,language_id)
ivg_embedding=get_language_embedding(ivg_files,path_to_wavs,language_id)
# Finally, you save the embedding in the computer
torch.save(at_embedding,'at_emb_trained.pt')
torch.save(vd_embedding,'vd_emb_trained.pt')
torch.save(goi_embedding,'goi_emb_trained.pt')
torch.save(ivg_embedding,'ivg_emb_trained.pt')
# torch.save(sp_embedding,'fr_emb.pt')
# for j,wav in enumerate(wavs):
# if j>99:
# break
# spkr_names.append(dir)
# emb = generate_embedding(wav)
# print(emb.shape)
# temp.append(emb.squeeze().detach().cpu().numpy())
# all_embeddings.append(emb.squeeze().detach().cpu().numpy())
# centroid = np.stack(temp)
# print(centroid)
# centroid = np.mean(centroid,axis=0)
# print(centroid.shape)
# spkr_names.append(dir + '_centroid')
# # torch.save(centroid,os.path.join('./centroids_16k_denoised',dir+'.pt'))
# all_embeddings.append(centroid)
# print(len(all_embeddings))
# print(len(spkr_names))
# centroids = [i.detach().cpu().numpy() for i in centroids]
# print(centroids)
# at_embedding = torch.load('./embeds/at_emb.pt')
# vd_embedding = torch.load('./embeds/vd_emb.pt')
# goi_embedding = torch.load('./embeds/goi_emb.pt')
# ivg_embedding = torch.load('./embeds/ivg_emb.pt')
# print(at_embedding.shape)
# print(vd_embedding.shape)
# print(goi_embedding.shape)
# print(ivg_embedding.shape)
# print(type(at_embedding))
# print(type(vd_embedding))
# print(type(goi_embedding))
# print(type(ivg_embedding))
# centroids = []
# centroids.append(at_embedding)
# centroids.append(vd_embedding)
# centroids.append(goi_embedding)
# centroids.append(ivg_embedding)
# print(len(centroids))
# tsne = TSNE(n_components=2, verbose=1, perplexity=40, n_iter=300)
# transformed = tsne.fit_transform(centroids)
# data = {
# "dim-1": transformed[:, 0],
# "dim-2": transformed[:, 1],
# "label": ['at','vd','goi','ivg'],
# }
# plt.figure()
# sns.scatterplot(
# x="dim-1",
# y="dim-2",
# hue="label",
# palette=sns.color_palette(n_colors=121),
# data=data,
# legend="full",
# )
# plt.legend(loc="center left", bbox_to_anchor=(1, 0.5))
# plt.tight_layout()
# plt.savefig('imagen_centroids.png')
# print(all_embeddings) |