def main(config, random_seed, dist, apply_normalization, n_pad):
"""
构建 IRM（Ideal ratio mask）语音增强数据集
数据集为语句级别，带噪语音和它相应纯净语音的频谱尺寸相同
Steps:
1. 加载纯净语音信号
2. 加载噪声文件
3. 在纯净语音信号上叠加噪声信号
4. 计算频谱，mask等
5. 分别存储带噪语音的频谱与 mask
Args:
config (dict): 配置信息
random_seed (int): 随机种子
dist (str): 输出结果的目录
apply_normalization (bool): 是否对 mixture 语音进行规范化
n_pad (int): mixture 语音中帧的拓展范围，拓展后中心帧对应 mask 中的一帧
Dataset:
dataset_1/
mixture.npy
mask.npy
...
mixture.npy is {
"0001_babble_-5": (257, T * (n_pad * 2 + 1)),
"0001_babble_-10": (257, T * T * (n_pad * 2 + 1))
...
}
mask.npy is {
"0001_babble_-5": (257, T),
"0001_babble_-10": (257, T),
...
}
"""
global clean_lps
np.random.seed(random_seed)
dist_dir = Path(dist)
# 以遍历的方式读取 config.json 中各个数据集的配置项
for dataset_itx, dataset_cfg in enumerate(config["dataset"], start=1):
dataset_dir = dist_dir / dataset_cfg["name"]
prepare_empty_dirs([dataset_dir])
print("=" * 12 +
f"Building set {dataset_itx}: {dataset_cfg['name']} set" +
"=" * 12)
# 加载纯净语音信号，存至 list 中
clean_cfg = dataset_cfg["clean"]
clean_speech_paths = librosa.util.find_files(
directory=clean_cfg["database"],
ext=clean_cfg["ext"],
recurse=clean_cfg["recurse"],
limit=clean_cfg["limit"],
offset=clean_cfg["offset"])
random.shuffle(clean_speech_paths)
clean_ys = load_wavs(
file_paths=clean_speech_paths,
sr=clean_cfg["sampling_rate"],
min_sampling=clean_cfg["min_sampling"],
)
print("Loaded clean speeches.")
# 加载噪声信号，存至 dict 中
noise_cfg = dataset_cfg["noise"]
noise_database_dir = Path(noise_cfg["database"])
noise_ys = {}
for noise_type in tqdm(noise_cfg["types"], desc="Loading noise files"):
mixture, _ = librosa.load(
(noise_database_dir / (noise_type + ".wav")).as_posix(),
sr=noise_cfg["sampling_rate"])
noise_ys[noise_type] = mixture
print("Loaded noise.")
# 合成带噪语音
mixture_store = {}
mask_store = {}
for i, clean in tqdm(enumerate(clean_ys, start=1), desc="合成带噪语音"):
num = str(i).zfill(4)
for snr in dataset_cfg["snr"]:
for noise_type in noise_ys.keys():
basename_text = f"{num}_{noise_type}_{snr}"
clean, noise = corrected_the_length_of_noise_and_clean_speech(
clean_y=clean, noise_y=noise_ys[noise_type])
mixture = add_noise_for_waveform(clean, noise, int(snr))
mixture_mag = mag(mixture)
clean_mag = mag(clean)
noise_mag = mag(noise)
if apply_normalization:
mixture_mag = input_normalization(mixture_mag)
mixture_mag = unfold_spectrum(mixture_mag, n_pad=n_pad)
mask = noise_mag / (noise_mag + clean_mag)
assert mixture_mag.shape[0] == mask.shape[0] == 257
mixture_store[basename_text] = mixture_mag
mask_store[basename_text] = mask
print(f"Synthesize finished，storing file...")
joblib.dump(mask_store, (dataset_dir / "mask.pkl").as_posix())
joblib.dump(mixture_store, (dataset_dir / "mixture.pkl").as_posix())
def main(config, random_seed, dist, n_pad):
"""
构建*频域*上的语音增强数据集（Log Power Spectrum）
每句带噪语音的时间步上都包含多帧，多帧的中心帧对应这个时间步上的一帧纯净语音
中心帧前面的时间帧：
中心帧后面的时间帧：
TODO 文档等待进一步更新
Steps:
1. 加载纯净语音信号
2. 加载噪声文件
3. 在纯净语音信号上叠加噪声信号
4. 分别计算 LPS 特征
5. 将带噪语音的 LPS 特征进行拓展
5. 分别存储带噪语音与纯净语音
Args:
config (dict): 配置信息
random_seed (int): 随机种子
dist (str): 输出结果的目录
n_pad (int): 带噪语音的拓展大小
Dataset:
dataset_1/
mixture.npy
clean.npy
...
mixture.npy is {
"0001_babble_-5": (257 * 3 * , T),
"0001_babble_-10": (257 * 3, T),
...
}
clean.npy is {
"0001": (257, T),
"0002": (257, T),
...
}
"""
global clean_lps
np.random.seed(random_seed)
dist_dir = Path(dist)
# 以遍历的方式读取 config.json 中各个数据集的配置项
for dataset_itx, dataset_cfg in enumerate(config["dataset"], start=1):
dataset_dir = dist_dir / dataset_cfg["name"]
prepare_empty_dirs([dataset_dir])
print("=" * 12 +
f"Building set {dataset_itx}: {dataset_cfg['name']} set" +
"=" * 12)
# 加载纯净语音信号，存至 list 中
clean_cfg = dataset_cfg["clean"]
clean_speech_paths = librosa.util.find_files(
directory=clean_cfg["database"],
ext=clean_cfg["ext"],
recurse=clean_cfg["recurse"],
limit=clean_cfg["limit"],
offset=clean_cfg["offset"])
random.shuffle(clean_speech_paths)
clean_ys = load_wavs(
file_paths=clean_speech_paths,
sr=clean_cfg["sampling_rate"],
min_sampling=clean_cfg["min_sampling"],
)
print("Loaded clean speeches.")
# 加载噪声信号，存至 dict 中
noise_cfg = dataset_cfg["noise"]
noise_database_dir = Path(noise_cfg["database"])
noise_ys = {}
for noise_type in tqdm(noise_cfg["types"], desc="Loading noise files"):
mixture, _ = librosa.load(
(noise_database_dir / (noise_type + ".wav")).as_posix(),
sr=noise_cfg["sampling_rate"])
noise_ys[noise_type] = mixture
print("Loaded noise.")
# 合成带噪语音
mixture_store = {}
clean_store = {}
for i, clean in tqdm(enumerate(clean_ys, start=1), desc="合成带噪语音"):
num = str(i).zfill(4)
for snr in dataset_cfg["snr"]:
for noise_type in noise_ys.keys():
basename_text = f"{num}_{noise_type}_{snr}"
clean, noise = corrected_the_length_of_noise_and_clean_speech(
clean_y=clean, noise_y=noise_ys[noise_type])
mixture = add_noise_for_waveform(clean, noise, int(snr))
assert len(mixture) == len(clean) == len(noise)
mixture_lps = lps(mixture)
clean_lps = lps(clean)
mixture_lps = unfold_spectrum(mixture_lps, n_pad=n_pad)
assert mixture_lps.shape[0] == clean_lps.shape[0] == 257
mixture_store[basename_text] = mixture_lps
clean_store[num] = clean_lps
print(f"Synthesize finished，storing file...")
joblib.dump(clean_store, (dataset_dir / "clean.pkl").as_posix())
joblib.dump(mixture_store, (dataset_dir / "mixture.pkl").as_posix())
def main(config, random_seed, dist):
"""
构建时域上的语音增强数据集
Steps:
1. 加载纯净语音信号
2. 加载噪声文件
3. 在纯净语音信号上叠加噪声信号
4. 分别存储带噪语音与纯净语音
Args:
config (dict): 配置信息
random_seed (int): 随机种子
dist (str): 输出结果的目录
Dataset:
dataset_1/
mixture.npy
clean.npy
...
mixture.npy is {
"0001_babble_-5": [signals, ...],
"0001_babble_-10": [signals, ...],
...
}
clean.npy is {
"0001": [signals, ...],
"0002": [signals, ...],
...
}
"""
np.random.seed(random_seed)
dist_dir = Path(dist)
# 以遍历的方式读取 config.json 中各个数据集的配置项
for dataset_itx, dataset_cfg in enumerate(config["dataset"], start=1):
dataset_dir = dist_dir / dataset_cfg["name"]
prepare_empty_dirs([dataset_dir, dataset_dir / "Clean", dataset_dir / "Noisy"])
print("=" * 12 + f"Building set {dataset_itx}: {dataset_cfg['name']} set" + "=" * 12)
# 加载纯净语音信号，存至 list 中
clean_cfg = dataset_cfg["clean"]
clean_speech_paths = librosa.util.find_files(
directory=clean_cfg["database"],
ext=clean_cfg["ext"],
recurse=clean_cfg["recurse"],
limit=clean_cfg["limit"],
offset=clean_cfg["offset"]
)
random.shuffle(clean_speech_paths)
clean_ys = load_wavs(
file_paths=clean_speech_paths,
sr=clean_cfg["sampling_rate"],
min_sampling=clean_cfg["min_sampling"],
)
print("Loaded clean speeches.")
# 加载噪声信号，存至 dict 中
noise_cfg = dataset_cfg["noise"]
noise_database_dir = Path(noise_cfg["database"])
noise_ys = {}
for noise_type in tqdm(noise_cfg["types"], desc="Loading noise files"):
mixture, _ = librosa.load(
(noise_database_dir / (noise_type + ".wav")).as_posix(),
sr=noise_cfg["sampling_rate"])
noise_ys[noise_type] = mixture
print("Loaded noise.")
# 合成带噪语音
n = 0
for i, clean in tqdm(enumerate(clean_ys, start=1), desc="合成带噪语音"):
for snr in dataset_cfg["snr"]:
for noise_type in noise_ys.keys():
clean, noise = corrected_the_length_of_noise_and_clean_speech(
clean_y=clean,
noise_y=noise_ys[noise_type]
)
mixture = add_noise_for_waveform(clean, noise, int(snr))
assert len(mixture) == len(clean) == len(noise)
fname = f"{dataset_cfg['name']}_{n}.wav"
librosa.output.write_wav((dataset_dir / "Clean" / fname).as_posix(), clean, sr=16000)
librosa.output.write_wav((dataset_dir / "Noisy" / fname).as_posix(), mixture, sr=16000)
n += 1
def main(config):
OUTPUT_DIR = Path(config["output_dir"])
SAMPLING_RATE = config["sampling_rate"]
for j, dataset_cfg in enumerate(config["datasets"]):
print(f"============ Building set {j + 1}: {dataset_cfg['name']} set ============")
dataset_dir: Path = OUTPUT_DIR / dataset_cfg["name"]
prepare_empty_dirs([dataset_dir])
"""============ clean speeches ============"""
clean_meta = dataset_cfg["clean"]
clean_speech_paths = librosa.util.find_files(
directory=clean_meta["database"],
ext=clean_meta["ext"],
recurse=clean_meta["recurse"],
limit=None,
offset=clean_meta["offset"]
)
random.shuffle(clean_speech_paths)
# 加载纯净语音时可以指定 minimum_sampling 参数，控制加载语音需要满足的最小采样点数
# 但在加载噪声时则没有这个参数。如果在合成带噪语音阶段发现噪声长度小于语音长度，则将噪声复制多次再合成带噪语音。
clean_ys = load_wavs(
file_paths=clean_speech_paths,
limit=clean_meta["limit"],
sr=SAMPLING_RATE,
minimum_sampling=clean_meta["minimum_sampling"],
)
print("Loaded clean speeches.")
"""============ noise speeches ============"""
noise_meta = dataset_cfg["noise"]
noise_database_dir = Path(noise_meta["database"])
noise_ys = {}
for noise_type in tqdm(noise_meta["types"], desc="Loading noise files"):
noise_y, _ = librosa.load((noise_database_dir / (noise_type + ".wav")).as_posix(), sr=SAMPLING_RATE)
noise_ys[noise_type] = noise_y
print("Loaded noise.")
"""============ 合成 ============"""
# 带噪
for i, SNR in enumerate(dataset_cfg["SNRs"]):
store = {}
clean_store = {}
for j, clean_y in tqdm(enumerate(clean_ys, 1), desc="Add noise for clean waveform"):
for noise_type in noise_ys.keys():
output_wav_basename_text = f"{str(j).zfill(4)}_{noise_type}"
clean_y, noise_y = corrected_length(
clean_y=clean_y,
noise_y=noise_ys[noise_type]
)
noisy_y = add_noise_for_waveform(clean_y, noise_y, int(SNR))
assert len(noisy_y) == len(clean_y) == len(noise_y)
"""
SNR == -5 是整个模型的输入，使用 7 帧
剩余的信噪比和纯净语音为模型训练的目标，使用 1 帧
"""
if SNR == -5:
tmp_lps = torch.Tensor(lps(noisy_y, pad=3).T).unfold(0, 7, 1)
store[output_wav_basename_text] = tmp_lps.reshape(tmp_lps.shape[0], -1).numpy()
else:
store[output_wav_basename_text] = lps(noisy_y).T
if i == 0:
clean_store[output_wav_basename_text] = lps(clean_y).T
print(f"Synthesize dB{SNR} finished，storing NPY file...")
if clean_store:
print("Saving clean NPY file...")
np.save((dataset_dir / "clean.npy").as_posix(), clean_store)
np.save((dataset_dir / f"dB{SNR}.npy").as_posix(), store)
def main(config, random_seed, dist):
"""
构建时域上的语音增强数据集
Steps:
1. 加载纯净语音信号
2. 加载噪声文件
3. 在纯净语音信号上叠加噪声信号
4. 分别存储带噪语音与纯净语音
Args:
config (dict): 配置信息
random_seed (int): 随机种子
dist (str): 输出结果的目录
Dataset:
dataset_1/
mixture.npy
clean.npy
...
mixture.npy is {
"0001_babble_-5": [signals, ...],
"0001_babble_-10": [signals, ...],
...
}
clean.npy is {
"0001": [signals, ...],
"0002": [signals, ...],
...
}
"""
np.random.seed(random_seed)
dist_dir = Path(dist)
# 以遍历的方式读取 config.json 中各个数据集的配置项
for dataset_itx, dataset_cfg in enumerate(config["dataset"], start=1):
dataset_dir = dist_dir / dataset_cfg["name"]
prepare_empty_dirs([dataset_dir])
print("=" * 12 +
f"Building set {dataset_itx}: {dataset_cfg['name']} set" +
"=" * 12)
# 加载纯净语音信号，存至 list 中
clean_cfg = dataset_cfg["clean"]
clean_speech_paths = librosa.util.find_files(
directory=clean_cfg["database"],
ext=clean_cfg["ext"],
recurse=clean_cfg["recurse"],
limit=clean_cfg["limit"],
offset=clean_cfg["offset"])
random.shuffle(clean_speech_paths)
clean_ys = load_wavs(
file_paths=clean_speech_paths,
sr=clean_cfg["sampling_rate"],
min_sampling=clean_cfg["min_sampling"],
)
print("Loaded clean speeches.")
# 加载噪声信号，存至 dict 中
noise_cfg = dataset_cfg["noise"]
noise_database_dir = Path(noise_cfg["database"])
noise_ys = {}
for noise_type in tqdm(noise_cfg["types"], desc="Loading noise files"):
mixture, _ = librosa.load(
(noise_database_dir / (noise_type + ".wav")).as_posix(),
sr=noise_cfg["sampling_rate"])
noise_ys[noise_type] = mixture
print("Loaded noise.")
# 合成带噪语音
mixture_store = {}
clean_store = {}
for i, clean in tqdm(enumerate(clean_ys, start=1), desc="合成带噪语音"):
num = str(i).zfill(4)
for snr in dataset_cfg["snr"]:
for noise_type in noise_ys.keys():
basename_text = f"{num}_{noise_type}_{snr}"
clean, noise = corrected_the_length_of_noise_and_clean_speech(
clean_y=clean, noise_y=noise_ys[noise_type])
mixture = add_noise_for_waveform(clean, noise, int(snr))
assert len(mixture) == len(clean) == len(noise)
mixture_store[basename_text] = mixture
# 基于一条纯净语音可以合成多种类型的带噪语音，但仅存储一份纯净语音
clean_store[num] = clean
print(f"Synthesize finished，storing file...")
joblib.dump(clean_store, (dataset_dir / "clean.pkl").as_posix())
joblib.dump(mixture_store, (dataset_dir / "mixture.pkl").as_posix())