Spaces:
Build error
Build error
File size: 3,653 Bytes
fc8786d |
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 |
import soundfile as sf
import torch
import fire
import torch.nn.functional as F
from torchaudio.transforms import Resample
from models.ecapa_tdnn import ECAPA_TDNN_SMALL
MODEL_LIST = ['ecapa_tdnn', 'hubert_large', 'wav2vec2_xlsr', 'unispeech_sat', "wavlm_base_plus", "wavlm_large"]
def init_model(model_name, checkpoint=None):
if model_name == 'unispeech_sat':
config_path = 'config/unispeech_sat.th'
model = ECAPA_TDNN_SMALL(feat_dim=1024, feat_type='unispeech_sat', config_path=config_path)
elif model_name == 'wavlm_base_plus':
config_path = None
model = ECAPA_TDNN_SMALL(feat_dim=768, feat_type='wavlm_base_plus', config_path=config_path)
elif model_name == 'wavlm_large':
config_path = None
model = ECAPA_TDNN_SMALL(feat_dim=1024, feat_type='wavlm_large', config_path=config_path)
elif model_name == 'hubert_large':
config_path = None
model = ECAPA_TDNN_SMALL(feat_dim=1024, feat_type='hubert_large_ll60k', config_path=config_path)
elif model_name == 'wav2vec2_xlsr':
config_path = None
model = ECAPA_TDNN_SMALL(feat_dim=1024, feat_type='wav2vec2_xlsr', config_path=config_path)
else:
model = ECAPA_TDNN_SMALL(feat_dim=40, feat_type='fbank')
if checkpoint is not None:
state_dict = torch.load(checkpoint, map_location=lambda storage, loc: storage)
model.load_state_dict(state_dict['model'], strict=False)
return model
def verification(model_name, wav1, wav2, use_gpu=True, checkpoint=None):
assert model_name in MODEL_LIST, 'The model_name should be in {}'.format(MODEL_LIST)
model = init_model(model_name, checkpoint)
wav1, sr1 = sf.read(wav1)
wav2, sr2 = sf.read(wav2)
wav1 = torch.from_numpy(wav1).unsqueeze(0).float()
wav2 = torch.from_numpy(wav2).unsqueeze(0).float()
resample1 = Resample(orig_freq=sr1, new_freq=16000)
resample2 = Resample(orig_freq=sr2, new_freq=16000)
wav1 = resample1(wav1)
wav2 = resample2(wav2)
if use_gpu:
model = model.cuda()
wav1 = wav1.cuda()
wav2 = wav2.cuda()
model.eval()
with torch.no_grad():
emb1 = model(wav1)
emb2 = model(wav2)
sim = F.cosine_similarity(emb1, emb2)
# print("The similarity score between two audios is {:.4f} (-1.0, 1.0).".format(sim[0].item()))
return sim[0].item()
def verification_batch(model_name, batch_wav1, batch_wav2, use_gpu=True, checkpoint=None):
assert model_name in MODEL_LIST, 'The model_name should be in {}'.format(MODEL_LIST)
model = init_model(model_name, checkpoint)
# print(str(batch_wav1[0]))
sr1 = sf.read(str(batch_wav1[0]))[1]
sr2 = sf.read(str(batch_wav2[0]))[1]
# print(sr1)
batch_wav1 = [torch.from_numpy(sf.read(wav)[0][:50000]).unsqueeze(0).float() for wav in batch_wav1]
batch_wav2 = [torch.from_numpy(sf.read(wav)[0][:50000]).unsqueeze(0).float() for wav in batch_wav2]
resample1 = Resample(orig_freq=sr1, new_freq=16000)
resample2 = Resample(orig_freq=sr2, new_freq=16000)
batch_wav1 = torch.cat([resample1(wav) for wav in batch_wav1], 0)
batch_wav2 = torch.cat([resample2(wav) for wav in batch_wav2], 0)
# print(batch_wav1.shape)
# print(batch_wav2.shape)
if use_gpu:
model = model.cuda()
batch_wav1 = batch_wav1.cuda()
batch_wav2 = batch_wav2.cuda()
model.eval()
with torch.no_grad():
emb1 = model(batch_wav1)
emb2 = model(batch_wav2)
sim = F.cosine_similarity(emb1, emb2 ,dim=-1)
return sim.cpu().numpy()
if __name__ == "__main__":
fire.Fire(verification)
|