def gen_post(feat_list, stat_file, model, win_size_before = 15, win_size_after = 15, num_targets = 31):
model.eval() # Put the model in test mode (the opposite of model.train(), essentially)
m, v = read_mv(stat_file)
if m is None or v is None:
raise Exception("mean or variance vector does not exist")
with open(feat_list) as f:
for line in f:
line = line.strip()
if len(line) < 1: continue
print ("generating features for file", line)
io = htk_io.fopen(line)
utt_feat = io.getall()
utt_feat -= m # normalize mean
utt_feat /= (np.sqrt(v) + eps) # normalize var
feat_numpy = org_data(utt_feat, win_size_before, win_size_after)
out_feat = np.zeros((utt_feat.shape[0], num_targets))
for i in range(feat_numpy.shape[0] // 100): # chop the speech into shorter segments, to prevent gpu out of memory
start_idx = i * 100
end_idx = i * 100 + 100
feat_chunk = feat_numpy[start_idx:end_idx]
feat_tensor = torch.from_numpy(feat_chunk).type(gpu_dtype)
x = Variable(feat_tensor.type(gpu_dtype), volatile = True)
scores = model(x)
out_feat[start_idx:end_idx] = scores.data.cpu().numpy()
num_remain = feat_numpy.shape[0] % 100
if num_remain > 0:
feat_chunk = feat_numpy[-num_remain:]
feat_tensor = torch.from_numpy(feat_chunk).type(gpu_dtype)
x = Variable(feat_tensor.type(gpu_dtype), volatile = True)
scores = model(x)
out_feat[-num_remain:] = scores.data.cpu().numpy()
out_feat = dct(out_feat, type=2, axis=1, norm='ortho')[:,1:numcep+1]
out_feat_delta = delta(out_feat, 2)
out_feat_ddelta = delta(out_feat_delta, 2)
out_feat = np.concatenate((out_feat, out_feat_delta, out_feat_ddelta), axis = 1)
out_file = line.replace(".fea", ".mfc")
io = htk_io.fopen(out_file, mode="wb", veclen = out_feat.shape[1])
io.writeall(out_feat)
print ("features saved in %s\n" %out_file)
评论列表
文章目录
