neverlib 0.2.2__py3-none-any.whl → 0.2.4__py3-none-any.whl

neverlib/.history/Docs/audio_aug/test_snr_20250806011331.py

@@ -0,0 +1,75 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-03-24 10:00:14
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/1/29
+"""
+snr增强, 对安静的测试集加指定snr的白噪
+"""
+import numpy as np
+import random
+import soundfile as sf
+from neverlib.metrics.snr import get_snr
+from neverlib.audio_aug import snr_aug_changeNoise, snr_aug_changeNoise_v2
+from neverlib.utils import EPS
+
+
+def add_noise(clean, snr_range, snr_aug_rate):
+    """
+    Args:
+        clean: (*, C)
+        snr_range: snr范围 [min, max]
+        snr_aug_rate: snr增强率
+    Returns:
+    """
+    if random.random() < snr_aug_rate:
+        snr = random.uniform(snr_range[0], snr_range[1])
+        noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+        noisy = snr_aug_changeNoise(clean, noise, snr)
+    else:
+        noisy = clean
+        snr = 100
+    return noisy, snr
+
+
+def snr_aug1(clean, snr):
+    """ 白噪
+    Args:
+        clean: (*, C)
+        snr: snr值
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+
+    noisy, _ = snr_aug_changeNoise(clean, noise, snr)
+
+    return noisy
+
+
+def snr_aug2(clean, snr):
+    """
+    Args:
+        clean: (*, C)
+        snr_range: snr范围 [min, max]
+        snr_aug_rate: snr增强率
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+    noisy, _ = snr_aug_changeNoise_v2(clean, noise, snr)
+    return noisy
+
+
+if __name__ == "__main__":
+    clean_path = "../wav_data/TIMIT.wav"
+    clean, fs = sf.read(clean_path, always_2d=True, dtype="float32")
+
+    noisy1 = snr_aug1(clean, 10)
+    noisy2 = snr_aug2(clean, 10)
+    print(clean.shape)
+    print(noisy1.shape, noisy2.shape)
+
+    print(get_snr(clean, noisy1 - clean))  # 10.000000000480982
+    print(get_snr(clean, noisy2 - clean))  # 9.999870642663442

neverlib/.history/Docs/audio_aug/test_snr_20250806011342.py

@@ -0,0 +1,57 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-03-24 10:00:14
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/1/29
+"""
+snr增强, 对安静的测试集加指定snr的白噪
+"""
+import numpy as np
+import random
+import soundfile as sf
+from neverlib.metrics.snr import get_snr
+from neverlib.audio_aug import snr_aug_changeNoise, snr_aug_changeNoise_v2
+from neverlib.utils import EPS
+
+
+def snr_aug1(clean, snr):
+    """ 白噪
+    Args:
+        clean: (*, C)
+        snr: snr值
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+
+    noisy, _ = snr_aug_changeNoise(clean, noise, snr)
+
+    return noisy
+
+
+def snr_aug2(clean, snr):
+    """
+    Args:
+        clean: (*, C)
+        snr_range: snr范围 [min, max]
+        snr_aug_rate: snr增强率
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+    noisy, _ = snr_aug_changeNoise_v2(clean, noise, snr)
+    return noisy
+
+
+if __name__ == "__main__":
+    clean_path = "../wav_data/TIMIT.wav"
+    clean, fs = sf.read(clean_path, always_2d=True, dtype="float32")
+
+    noisy1 = snr_aug1(clean, 10)
+    noisy2 = snr_aug2(clean, 10)
+    print(clean.shape)
+    print(noisy1.shape, noisy2.shape)
+
+    print(get_snr(clean, noisy1 - clean))  # 10.000000000480982
+    print(get_snr(clean, noisy2 - clean))  # 9.999870642663442

neverlib/.history/Docs/audio_aug/test_snr_20250806011352.py

@@ -0,0 +1,57 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-03-24 10:00:14
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/1/29
+"""
+snr增强, 对安静的测试集加指定snr的白噪
+"""
+import numpy as np
+import random
+import soundfile as sf
+from neverlib.metrics.snr import get_snr
+from neverlib.audio_aug import snr_aug_changeNoise, snr_aug_changeNoise_v2
+from neverlib.utils import EPS
+
+
+def snr_aug1(clean, snr):
+    """ 白噪
+    Args:
+        clean: (*, C)
+        snr: snr值
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+
+    noisy, _ = snr_aug_changeNoise(clean, noise, snr)
+
+    return noisy
+
+
+def snr_aug2(clean, snr):
+    """
+    Args:
+        clean: (*, C)
+        snr_range: snr范围 [min, max]
+        snr_aug_rate: snr增强率
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+    noisy, _ = snr_aug_changeNoise_v2(clean, noise, snr)
+    return noisy
+
+
+if __name__ == "__main__":
+    clean_path = "../../wav_data/TIMIT.wav"
+    clean, fs = sf.read(clean_path, always_2d=True, dtype="float32")
+
+    noisy1 = snr_aug1(clean, 10)
+    noisy2 = snr_aug2(clean, 10)
+    print(clean.shape)
+    print(noisy1.shape, noisy2.shape)
+
+    print(get_snr(clean, noisy1 - clean))  # 10.000000000480982
+    print(get_snr(clean, noisy2 - clean))  # 9.999870642663442

neverlib/.history/Docs/audio_aug/test_snr_20250806011403.py

@@ -0,0 +1,57 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-03-24 10:00:14
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/1/29
+"""
+snr增强, 对安静的测试集加指定snr的白噪
+"""
+import numpy as np
+import random
+import soundfile as sf
+from neverlib.metrics.snr import get_snr
+from neverlib.audio_aug import snr_aug_changeNoise, snr_aug_changeNoise_v2
+from neverlib.utils import EPS
+
+
+def snr_aug1(clean, snr):
+    """ 白噪
+    Args:
+        clean: (*, C)
+        snr: snr值
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+
+    noisy, _ = snr_aug_changeNoise(clean, noise, snr)
+
+    return noisy
+
+
+def snr_aug2(clean, snr):
+    """
+    Args:
+        clean: (*, C)
+        snr_range: snr范围 [min, max]
+        snr_aug_rate: snr增强率
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+    noisy, _ = snr_aug_changeNoise_v2(clean, noise, snr)
+    return noisy
+
+
+if __name__ == "__main__":
+    clean_path = "../../data/TIMIT.wav"
+    clean, fs = sf.read(clean_path, always_2d=True, dtype="float32")
+
+    noisy1 = snr_aug1(clean, 10)
+    noisy2 = snr_aug2(clean, 10)
+    print(clean.shape)
+    print(noisy1.shape, noisy2.shape)
+
+    print(get_snr(clean, noisy1 - clean))  # 10.000000000480982
+    print(get_snr(clean, noisy2 - clean))  # 9.999870642663442

neverlib/.history/Docs/audio_aug/test_snr_20250806011413.py

@@ -0,0 +1,57 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-03-24 10:00:14
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/1/29
+"""
+snr增强, 对安静的测试集加指定snr的白噪
+"""
+import numpy as np
+import random
+import soundfile as sf
+from neverlib.metrics.snr import get_snr
+from neverlib.audio_aug import snr_aug_changeNoise, snr_aug_changeNoise_v2
+from neverlib.utils import EPS
+
+
+def snr_aug1(clean, snr):
+    """ 白噪
+    Args:
+        clean: (*, C)
+        snr: snr值
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+
+    noisy, _ = snr_aug_changeNoise(clean, noise, snr)
+
+    return noisy
+
+
+def snr_aug2(clean, snr):
+    """
+    Args:
+        clean: (*, C)
+        snr_range: snr范围 [min, max]
+        snr_aug_rate: snr增强率
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+    noisy, _ = snr_aug_changeNoise_v2(clean, noise, snr)
+    return noisy
+
+
+if __name__ == "__main__":
+    clean_path = "../../data/white.wav"
+    clean, fs = sf.read(clean_path, always_2d=True, dtype="float32")
+
+    noisy1 = snr_aug1(clean, 10)
+    noisy2 = snr_aug2(clean, 10)
+    print(clean.shape)
+    print(noisy1.shape, noisy2.shape)
+
+    print(get_snr(clean, noisy1 - clean))  # 10.000000000480982
+    print(get_snr(clean, noisy2 - clean))  # 9.999870642663442

neverlib/.history/Docs/audio_aug/test_snr_20250806011435.py

@@ -0,0 +1,55 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-03-24 10:00:14
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/1/29
+"""
+snr增强, 对安静的测试集加指定snr的白噪
+"""
+import numpy as np
+import random
+import soundfile as sf
+from neverlib.metrics.snr import get_snr
+from neverlib.audio_aug import snr_aug_changeNoise, snr_aug_changeNoise_v2
+from neverlib.utils import EPS
+
+
+def snr_aug1(clean, snr):
+    """ 白噪
+    Args:
+        clean: (*, C)
+        snr: snr值
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+
+    noisy, _ = snr_aug_changeNoise(clean, noise, snr)
+
+    return noisy
+
+
+def snr_aug2(clean, snr):
+    """
+    Args:
+        clean: (*, C)
+        snr_range: snr范围 [min, max]
+        snr_aug_rate: snr增强率
+    Returns:
+    """
+    noise = np.random.randn(*clean.shape)  # 生成和clean等长的白噪
+    noisy, _ = snr_aug_changeNoise_v2(clean, noise, snr)
+    return noisy
+
+
+if __name__ == "__main__":
+    clean_path = "../../data/white.wav"
+    clean, fs = sf.read(clean_path, always_2d=True, dtype="float32")
+
+    noisy1 = snr_aug1(clean, 10)
+    noisy2 = snr_aug2(clean, 10)
+
+    print(get_snr(clean, noisy1 - clean))  # 10.000000000480982
+    print(get_snr(clean, noisy2 - clean))  # 9.999870642663442

neverlib/.history/Docs/vad/1_20250810032417.py

@@ -0,0 +1,39 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-08-05 23:42:06
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never.Ling
+# Date: 2022/8/2
+"""
+案例来源：https://github.com/snakers4/silero-vad
+API文档：https://github.com/snakers4/silero-vad/blob/master/utils_vad.py
+"""
+from rVADfast import rVADfast
+import soundfile as sf 
+import numpy as np
+import matplotlib.pyplot as plt
+from neverlib.filter import HPFilter
+from neverlib.audio_aug import volume_norm
+
+
+sr = 16000
+vad = rVADfast()
+wav_path = "../../data/vad_example.wav"
+wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")   # (xxx, ch)
+assert wav_sr == sr, f"音频采样率为{wav_sr},期望{sr}"
+wav = HPFilter(wav, sr=16000, order=6, cutoff=100)
+wav = volume_norm(wav)
+
+vad_labels, vad_timestamps = vad(wav, wav_sr)
+print(len(wav))
+print(len(vad_labels))
+print(len(vad_timestamps))
+
+
+
+
+
+
+

neverlib/.history/audio_aug/audio_aug_20250806010451.py

@@ -0,0 +1,125 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/27
+"""
+
+"""
+import random
+import numpy as np
+import soundfile as sf
+from scipy import signal
+from neverlib.utils import EPS
+
+
+def volume_norm(wav):
+    """
+    音量归一化
+    :param wav: (T,)
+    :return: (T,)
+    """
+    wav = wav / (np.max(np.abs(wav)) + 1e-8)
+    return wav
+
+
+def add_reverb(wav, rir, ratio=1, mode="same"):
+    """添加混响,
+    Args:
+        wav: [T, channel]
+        rir: [T, channel]
+        ratio:  0-1
+        mode: "same" for SE or "full" for kws
+    """
+    if random.random() < ratio:
+        wav = signal.fftconvolve(wav, rir, mode=mode)  # (28671, 3)
+        # note: 建议过完添加混响后再进行归一化, 否则可能会出现溢出
+        # wav = volume_norm(wav)
+    return wav
+
+
+def snr_aug_changeNoise(clean, noise, snr):
+    """ 
+    保持语音不变, 改变噪声的幅度
+    snr = 10 * log10(signal_power / k*noise_power) 
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_power = np.mean(clean ** 2)  # 纯净语音功率
+    noise_power = np.mean(noise ** 2)  # 噪声功率
+    noise_scale = np.sqrt(clean_power / (noise_power * 10 ** (snr / 10) + EPS))
+    noisy = clean + noise_scale * noise
+    return noisy, noise_scale
+
+
+def snr_aug_changeClean(clean, noise, snr):
+    """ 
+    保持噪声不变, 改变语音的幅度
+    snr = 10 * log10(k*signal_power/ noise_power)
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_power = np.mean(clean ** 2)
+    noise_power = np.mean(noise ** 2)
+    clean_scale = np.sqrt(noise_power * 10 ** (snr / 10) / (clean_power + 1e-8))
+    noisy = clean * clean_scale + noise
+
+    return noisy, clean_scale
+
+
+def snr_aug_Interpolation(clean, noise, snr):
+    """
+    在已知clean_len<=noise_len的情况下
+    将clean插入到noise中的snr aug方法
+    Args:
+        clean: 语音
+        noise: 噪声
+        snr: snr=random.uniform(*snr_range)
+    """
+    clean_len, noise_len = clean.shape[0], noise.shape[0]
+    assert clean_len <= noise_len, f"clean_len must be less than noise_len."
+    noisy = noise.copy()
+    index = random.randint(0, noise_len - clean_len)
+    noise = noise[index:index + clean_len, :]
+    noisy_tmp, _ = snr_aug_changeClean(clean, noise, snr)
+    noisy[index:index + clean_len, :] = noisy_tmp
+    return noisy
+
+
+def get_audio_segments(wav_len, audio_path_list, sr=16000):
+    """
+    从音频列表中随机拼接指定长度音频
+    Args:
+        wav_len: 需要返回的音频长度
+        audio_path_list: 音频路径列表
+        sr: 采样率
+    Returns:返回指定长度的音频
+    """
+    audio_len = 0
+    wav_list = []
+    while audio_len < wav_len:
+        audio_path = random.choice(audio_path_list)
+        wav, wav_sr = sf.read(audio_path, always_2d=True, dtype='float32')
+        assert wav_sr == sr, f"音频采样率是{wav_sr}, 期望{sr}"
+        audio_len += len(wav)
+        wav_list.append(wav)
+    wav = np.concatenate(wav_list, axis=0)
+    if len(wav) > wav_len:
+        # 随机截取clean_len
+        start = random.randint(0, len(wav) - wav_len)
+        wav = wav[start:start + wav_len, :]
+    return wav
+
+
+def volume_aug(wav, range, rate, method="linmax"):
+    """音量增强 """
+    if random.random() < rate:
+        target_level = random.uniform(range[0], range[1])
+        if method == "dbrms":
+            wav_rms = (wav ** 2).mean() ** 0.5
+            scalar = 10 ** (target_level / 20) / (np.max(wav_rms) + EPS)
+        elif method == "linmax":
+            ipt_max = np.max(np.abs(wav))
+            # wav/wav_max*target_level=target_level_wav
+            # 处理后音频的 最大值就是target_level
+            scalar = target_level / (ipt_max + EPS)
+        else:
+            raise ValueError("method must be 'dbrms' or 'linmax'")
+        wav *= scalar
+    return wav

neverlib/.history/audio_aug/audio_aug_20250806010750.py

@@ -0,0 +1,138 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/27
+"""
+
+"""
+import random
+import numpy as np
+import soundfile as sf
+from scipy import signal
+from neverlib.utils import EPS
+
+
+def volume_norm(wav):
+    """
+    音量归一化
+    :param wav: (T,)
+    :return: (T,)
+    """
+    wav = wav / (np.max(np.abs(wav)) + 1e-8)
+    return wav
+
+
+def add_reverb(wav, rir, ratio=1, mode="same"):
+    """添加混响,
+    Args:
+        wav: [T, channel]
+        rir: [T, channel]
+        ratio:  0-1
+        mode: "same" for SE or "full" for kws
+    """
+    if random.random() < ratio:
+        wav = signal.fftconvolve(wav, rir, mode=mode)  # (28671, 3)
+        # note: 建议过完添加混响后再进行归一化, 否则可能会出现溢出
+        # wav = volume_norm(wav)
+    return wav
+
+
+def snr_aug_changeNoise(clean, noise, snr):
+    """ 
+    保持语音不变, 改变噪声的幅度
+    snr = 10 * log10(signal_power / k*noise_power) 
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_power = np.mean(clean ** 2)  # 纯净语音功率
+    noise_power = np.mean(noise ** 2)  # 噪声功率
+    noise_scale = np.sqrt(clean_power / (noise_power * 10 ** (snr / 10) + EPS))
+    noisy = clean + noise_scale * noise
+    return noisy, noise_scale
+
+
+def snr_aug_changeNoise(clean, noise, snr):
+    """ 
+    保持语音不变, 改变噪声的幅度
+    snr = 10 * log10(signal_power / k*noise_power) 
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_power = np.mean(clean ** 2)  # 纯净语音功率
+    noise_power = np.mean(noise ** 2)  # 噪声功率
+    noise_scale = np.sqrt(clean_power / (noise_power * 10 ** (snr / 10) + EPS))
+    noisy = clean + noise_scale * noise
+    return noisy, noise_scale
+
+
+def snr_aug_changeClean(clean, noise, snr):
+    """ 
+    保持噪声不变, 改变语音的幅度
+    snr = 10 * log10(k*signal_power/ noise_power)
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_power = np.mean(clean ** 2)
+    noise_power = np.mean(noise ** 2)
+    clean_scale = np.sqrt(noise_power * 10 ** (snr / 10) / (clean_power + 1e-8))
+    noisy = clean * clean_scale + noise
+
+    return noisy, clean_scale
+
+
+def snr_aug_Interpolation(clean, noise, snr):
+    """
+    在已知clean_len<=noise_len的情况下
+    将clean插入到noise中的snr aug方法
+    Args:
+        clean: 语音
+        noise: 噪声
+        snr: snr=random.uniform(*snr_range)
+    """
+    clean_len, noise_len = clean.shape[0], noise.shape[0]
+    assert clean_len <= noise_len, f"clean_len must be less than noise_len."
+    noisy = noise.copy()
+    index = random.randint(0, noise_len - clean_len)
+    noise = noise[index:index + clean_len, :]
+    noisy_tmp, _ = snr_aug_changeClean(clean, noise, snr)
+    noisy[index:index + clean_len, :] = noisy_tmp
+    return noisy
+
+
+def get_audio_segments(wav_len, audio_path_list, sr=16000):
+    """
+    从音频列表中随机拼接指定长度音频
+    Args:
+        wav_len: 需要返回的音频长度
+        audio_path_list: 音频路径列表
+        sr: 采样率
+    Returns:返回指定长度的音频
+    """
+    audio_len = 0
+    wav_list = []
+    while audio_len < wav_len:
+        audio_path = random.choice(audio_path_list)
+        wav, wav_sr = sf.read(audio_path, always_2d=True, dtype='float32')
+        assert wav_sr == sr, f"音频采样率是{wav_sr}, 期望{sr}"
+        audio_len += len(wav)
+        wav_list.append(wav)
+    wav = np.concatenate(wav_list, axis=0)
+    if len(wav) > wav_len:
+        # 随机截取clean_len
+        start = random.randint(0, len(wav) - wav_len)
+        wav = wav[start:start + wav_len, :]
+    return wav
+
+
+def volume_aug(wav, range, rate, method="linmax"):
+    """音量增强 """
+    if random.random() < rate:
+        target_level = random.uniform(range[0], range[1])
+        if method == "dbrms":
+            wav_rms = (wav ** 2).mean() ** 0.5
+            scalar = 10 ** (target_level / 20) / (np.max(wav_rms) + EPS)
+        elif method == "linmax":
+            ipt_max = np.max(np.abs(wav))
+            # wav/wav_max*target_level=target_level_wav
+            # 处理后音频的 最大值就是target_level
+            scalar = target_level / (ipt_max + EPS)
+        else:
+            raise ValueError("method must be 'dbrms' or 'linmax'")
+        wav *= scalar
+    return wav