def __getitem__(self, idx):
speaker = self.speakers[idx]
wav_files = glob.glob(speaker + '/*.WAV')
shuffle(wav_files)
wav_files = wav_files[0:self.utterance_number]
mel_dbs = []
for f in wav_files:
_, mel_db, _ = mfccs_and_spec(f, wav_process=True)
mel_dbs.append(mel_db)
return torch.Tensor(mel_dbs)
embedder_net = SpeechEmbedder()
embedder_net.load_state_dict(torch.load(model_path))
#speakers = ['付立', '刘子涵', '操镭', '李思琪', '李潇潇', '王润宇', '罗明双', '资礼波', '赵晴']
speakers = ['fuli', 'lzh', 'cl', 'lsq', 'lxx', 'wry', 'lms', 'zlb', 'zq']
enroll_wav_path = 'data_eng'
dict_spkid_embeddings = {}
enroll_nums = 5
for speaker in speakers:
#print(speaker)
speaker_wavs = glob.glob(os.path.join(enroll_wav_path, speaker, '*.wav'))
length = len(speaker_wavs)
print(speaker, speaker_wavs)
speaker_embeddings = np.zeros((length, 256),dtype=float)
for i in range(length):
#for wav in speaker_wavs:
embeddings = []
_, mel_db, _ = mfccs_and_spec(speaker_wavs[i], wav_process=True)
embeddings.append(mel_db)
embeddings = torch.Tensor(embeddings)
embeddings = torch.FloatTensor(np.transpose(embeddings, axes=(0,1,2)))
embeddings = embedder_net(embeddings)
embeddings = embeddings.cpu().detach().numpy()
speaker_embeddings[i] = embeddings
print(speaker_embeddings.shape)
dict_spkid_embeddings[speaker] = speaker_embeddings
np.save('enroll_9_.npy', dict_spkid_embeddings)
import time
import torch
import glob
import numpy as np
from torch.autograd import Variable
from sklearn.metrics.pairwise import cosine_similarity
from model import SpeechEmbedder
from utils import mfccs_and_spec
model_path = 'weights/epoch_3450_iteration_1487381_EER_0.05986112356185913.pth'
#data = './test_tisv/speaker2.npy'
data = 'data_change'
if (__name__=='__main__'):
print('Load model successfully.')
embedder_net = SpeechEmbedder()
embedder_net.load_state_dict(torch.load(model_path))
inputs_wavs = glob.glob(os.path.join(data, '*', '*.wav'))
for wav in inputs_wavs:
_, mel_db, _ = mfccs_and_spec(wav, wav_process=True)
inputs = torch.Tensor(mel_db)
print(inputs.shape)
inputs = inputs.unsqueeze(0)
inputs = torch.FloatTensor(np.transpose(inputs, axes=(0,1,2)))
outputs = embedder_net(inputs)
mel_db = outputs.detach().numpy()
output_path = wav[:-4]+'.npy'
np.save(output_path, mel_db)
total = 0
score = -10
name = 'Who?'
#speakers = ['付立', '刘子涵', '操镭', '李思琪', '李潇潇', '王润宇', '罗明双', '资礼波', '赵晴']
speakers = ['fuli', 'lzh', 'cl', 'lsq', 'lxx', 'wry', 'lms', 'zlb', 'zq']
data = 'media'
for speaker in speakers:
test_wavs = glob.glob(os.path.join(data, '{}_*.wav'.format(speaker)))
#print(test_wavs)
x = 0
y = 0
for test_wav in test_wavs:
scores = {}
test_mel = []
_, mel_db_test, _ = mfccs_and_spec(test_wav, wav_process=True)
test_mel.append(mel_db_test)
test_inputs = torch.Tensor(test_mel)
# print(test_inputs.size())
test_inputs = torch.FloatTensor(
np.transpose(test_inputs, axes=(0, 1, 2)))
#print(test_inputs.size())
test_output = embedder_net(test_inputs).cpu()
for speaker_name in enroll_centroid_embeddings.keys():
#print(enroll_centroid_embeddings[speaker_name].shape)
speaker_centroid_enroll = get_centroid(
enroll_centroid_embeddings[speaker_name],
len(enroll_centroid_embeddings[speaker_name]))
#print(speaker_centroid_enroll.shape)
score_speaker = cosine_similarity(
speaker_centroid_enroll.reshape(1, -1),