class TestSet(object):
"""docstring for TestSet."""
def __init__(self, datadir: str):
super(TestSet, self).__init__()
self.datadir = datadir
self.norm = Normalizer()
def choose(self):
'''choose one speaker for test'''
r = random.choice(speakers)
return r
def test_data(self, src_speaker=None):
'''choose one speaker for conversion'''
if src_speaker:
r_s = src_speaker
else:
r_s = self.choose()
p = os.path.join(self.datadir, r_s)
wavfiles = librosa.util.find_files(p, ext='wav')
res = {}
for f in wavfiles:
filename = os.path.basename(f)
wav, _ = librosa.load(f, sr=SAMPLE_RATE, dtype=np.float64)
f0, timeaxis, sp, ap, coded_sp = world_features(
wav, SAMPLE_RATE, FFTSIZE, FEATURE_DIM)
coded_sp_norm = self.norm.forward_process(coded_sp.T, r_s)
if not res.__contains__(filename):
res[filename] = {}
res[filename]['coded_sp_norm'] = np.asarray(coded_sp_norm)
res[filename]['f0'] = np.asarray(f0)
res[filename]['ap'] = np.asarray(ap)
return res, r_s
class TestSet(object):
def __init__(self, data_dir: str, sr: int):
super(TestSet, self).__init__()
self.data_dir = data_dir
self.norm = Normalizer()
self.sample_rate = sr
def choose(self):
r = choice(speakers)
return r
def test_data(self, src_speaker=None):
if src_speaker:
r_s = src_speaker
else:
r_s = self.choose()
p = os.path.join(self.data_dir, r_s)
wavfiles = librosa.util.find_files(p, ext='wav')
res = {}
for f in wavfiles:
filename = os.path.basename(f)
wav, _ = librosa.load(f, sr=self.sample_rate, dtype=np.float64)
f0, ap, mcep = cal_mcep(wav, self.sample_rate, FEATURE_DIM,
FFTSIZE, SHIFTMS, ALPHA)
mcep_norm = self.norm.forward_process(mcep, r_s)
if not res.__contains__(filename):
res[filename] = {}
res[filename]['mcep_norm'] = np.asarray(mcep_norm)
res[filename]['f0'] = np.asarray(f0)
res[filename]['ap'] = np.asarray(ap)
return res, r_s
class TestSet(object):
"""对于测试数据的的说明"""
def __init__(self, datadir: str):
super(TestSet, self).__init__()
self.datadir = datadir
# 定义对象norm方法为数据正则化方法
self.norm = Normalizer()
# 随机选取发音者方法
def choose(self):
"""为测试数据选择一个发音者"""
# 根据speakers这个序列中随机取出一个speaker作为目标发音者
r = random.choice(speakers)
return r
# 默认源发音者为空值
def test_data(self, src_speaker=None):
"""为转换数据选择一个发音者"""
# 如果传入了源发音者，即将这个参数赋值给r_s变量
if src_speaker:
r_s = src_speaker
# 如果没有参数传入
else:
# 就自动调用对象的choose方法随机选取一个发音者
r_s = self.choose()
# 将这个源发音者的名称和原本的数据集地址拼接在一起作为新路径
# 因为这个数据集是以发音者的名字作为子集的，所以要找到对应发音者的子数据集必须这样处理
# 如果采用不同的数据集命名方式那么处理方式就会不同
p = os.path.join(self.datadir, r_s)
# 根据路由找到该路径下的所有wav格式文件
wavfiles = librosa.util.find_files(p, ext='wav')
res = {}
# 遍历这个数据集对象
for f in wavfiles:
# 获取对应的文件名
filename = os.path.basename(f)
# librosa.load方法f为文件地址，sr参数为采样率，如果保存原有采样率则赋值为None
# dtype为精度，将返回一个音频时间序列和一个音频采样率
# 因为音频采样率不被使用，所以以_作为变量名占位保存
# 该方法还有mono ：bool，是否将信号转换为单声道
# offset ：float，在此时间之后开始阅读（以秒为单位）
# duration：float，仅加载这么多的音频（以秒为单位）
wav, _ = librosa.load(f, sr=SAMPLE_RATE, dtype=np.float64)
# 使用自定义的world_features获取对应的wav数据
f0, timeaxis, sp, ap, coded_sp = world_features(
wav, SAMPLE_RATE, FFTSIZE, FEATURE_DIM)
# 调用自定义的forward_process方法对编码频谱包络进行处理
coded_sp_norm = self.norm.forward_process(coded_sp.T, r_s)
# 将res添加对应的字典值，第一维键名为文件名
if not res.__contains__(filename):
res[filename] = {}
res[filename]['coded_sp_norm'] = np.asarray(coded_sp_norm)
res[filename]['f0'] = np.asarray(f0)
res[filename]['ap'] = np.asarray(ap)
# 返回处理后的数据字典与目标发音者标签
return res, r_s
