def mfcc_feature(self, file):
"""
mfcc 特征
:param file: (str, mandatory) 音频文件
:return: (array) mfcc 特征
"""
assert os.path.isfile(file), FileNotFoundException(file)
(sample, audio) = wav.read(file)
return mfcc(audio, sample, numcep=self.mfcc_dim)
def load(file):
"""
加载字典
:param file: (str, mandatory) 字典文件
:return: (dictionary) 字典对象
"""
assert os.path.isfile(file), FileNotFoundException(file)
with open(file, 'rb') as f_read:
audio_feature = pickle.loads(f_read.read())
return audio_feature
def spectrogram(self, file):
"""
获取声谱特征
:param file: (str, mandatory) 音频文件
:return: (array) 声谱特征。shape:[时序, frame_length // 2]
"""
assert os.path.isfile(file), FileNotFoundException(file)
# 1.提取信号。
# audio 为提取的数据。格式为[数据帧, 通道]。如果通道为0，则返回为1维度数据
# sample_rate 为采样率。
audio, sample_rate = sf.read(file, dtype='float32')
assert audio.ndim == 1, UnknownError("check input data")
# 2.分帧
# 2.1 根据帧长度、帧移、信号数据。判断最多可以获得多少帧数据。对多余数据抛弃。
# 例如信号数据 audio = [1,2,3,4,5,6,7,8,9]. 帧长度 = 4, 帧移 = 2
# 则第一时序数据帧是 [1,2,3,4], 第二时序数据帧是：[3,4,5,6], 第三时序数据帧是：[5,6,7,8]. 其中 9 这个数据点被抛弃。总共 3 个时序
trunc = (audio.shape[0] - self.frame_length) % self.frame_shift
audio = audio[:audio.shape[0] - trunc]
# 时序
sequential = (audio.shape[0] - self.frame_length) // self.frame_shift
assert sequential > 0, UnknownError(
"spectrogram feature sequence less than 0， "
"audio file:{}, audio shape:{}".format(file, audio.shape))
# 汉宁窗。shape: [frame_length]
window = np.hanning(self.frame_length)
# 3. 分帧 + 加窗 + 傅立叶变换
# 3.1 定义保存数据
feature = np.zeros((sequential, self.spectrogram_dim))
for i in range(sequential):
# 提取一个时序的帧数据
seq_frame = audio[i * self.frame_shift:i * self.frame_shift +
self.frame_length]
# 加窗
seq_frame = seq_frame * window
# 傅立叶变换.
seq_frame = np.fft.fft(seq_frame)
# 然后取数据的一半。因为数据是对称的
seq_frame = seq_frame[:self.spectrogram_dim]
# 取绝对值
feature[i] = np.abs(seq_frame)**2
feature = feature + 1.e-12
# assert feature.all() != 0, UnknownError("spectrogram feature cannot equal to 0, "
# "audio file:{}, audio shape:{}".format(file, audio.shape))
feature = np.log(feature)
return feature
def load(file):
"""
加载模型。
:param file: (str, mandatory) 模型文件
:return: (MFSSegmentation) 分词对象
"""
if not os.path.isfile(file):
raise FileNotFoundException(file)
with open(file, "rb") as f_read:
model = pickle.loads(f_read.read())
return model
def check_input_args(args):
""" 检查输入的参数 """
if args.audio is '':
print(Color.red("Please input audio file !"))
return False
else:
if not os.path.isfile(args.audio):
print(FileNotFoundException(args.audio))
return False
if args.top > args.beam:
print(Color.red("requested top:{} than the beam:{}".format(args.top, args.beam)))
return False
return True
def predict(self, audio, beam_width=3, top_paths=1, greedy=True):
""" 预测 """
assert os.path.isfile(audio), FileNotFoundException(audio)
features, input_length = self.audio_feature.feature(audio,
alignment=False)
features = features[np.newaxis, :, :]
features = tf.convert_to_tensor(features, dtype=tf.float32)
input_length = tf.reshape(tf.convert_to_tensor([input_length],
dtype=tf.int32),
shape=[1, 1])
sequence, _ = self.model.predict(features,
input_length,
beam_width=beam_width,
top_paths=top_paths,
greedy=greedy)
if greedy:
output = self.pinyin_dict.decoding(sequence)
else:
output = [self.pinyin_dict.decoding(seq) for seq in sequence]
return output
def load_dataset(source_data_path, data_path):
"""
加载数据集
:param data_path: (str, mandatory) 训练数据所在的路径
:param source_data_path: (str, mandatory) 实际数据路径
:return: (dict of list) 列表字典。返回音频文件路径，标签文件路径，标签数据路径，拼音字符，汉字
"""
audio_data, pinyin_data, chinese_data = list(), list(), list()
data_file = os.listdir(data_path)
for file in tqdm(data_file):
# 检查文件名后缀是不是 wav 音频类型文件。
if file[-3:] == 'wav':
# 标签文件。根据音频文件名得到标签文件名
label_file = file + '.trn'
# 判断标签文件是否存在。不存在则报错
assert label_file in data_file, FileNotFoundException(label_file)
# 音频文件路径和标签文件路径
audio_file_path = os.path.join(data_path, file)
label_file_path = os.path.join(data_path, label_file)
# 读取标签文件内容。找到对应标签数据路径
label_data_file_path = DataUtils.read_text_data(
label_file_path, show_progress_bar=False)
assert len(label_data_file_path) == 1, \
'Get redundant data path: {}, label_file_path:{}'.format(label_data_file_path, label_file_path)
# 重新拼接路径
label_data_file_path = os.path.join(
source_data_path, label_data_file_path[0].split('/')[-1])
assert os.path.isfile(label_data_file_path), FileNotFoundException(
label_data_file_path)
# 读取标签数据。包括字符、拼音
text_data = DataUtils.read_text_data(label_data_file_path,
handle_func=handle,
show_progress_bar=False)
chinese = text_data[0]
pinyin = text_data[1]
# 检查pinyin字符是否正确。主要判断每个拼音的最后一个字符是不是数字。不是数字则报错
for py in pinyin:
assert py[-1].isdigit(), "The last character:{} of Pinyin is not a number! " \
"pinyin_str:{}, " \
"label_data_file_path:{}, " \
"label_file_path:{}".format(py, pinyin, label_data_file_path,
label_file_path)
# 将由多个中文字符组成的词转换成单个字
new_chinese = list()
for ch in chinese:
new_chinese += list(ch)
chinese = new_chinese
# 检查是否是一个拼音对应中文字符， 如果不是则报错
# assert len(chinese) == len(pinyin), "The number of pinyin:{} and chinese:{} is different, " \
# "chinese:{}, pinyin:{}, file:{}".format(len(pinyin), len(chinese),
# chinese, pinyin,
# label_data_file_path)
audio_data.append(audio_file_path)
pinyin_data.append(pinyin)
chinese_data.append(chinese)
return audio_data, pinyin_data, chinese_data