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

neverlib/.history/metrics/spec_metric_20250816135537.py

@@ -0,0 +1,63 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-08-16 13:51:57
+Description: 
+'''
+
+import librosa
+import numpy as np
+import soundfile as sf
+
+def lsd(reference, estimate, n_fft=2048, hop_length=512, win_length=None):
+    """
+    计算两个一维音频信号之间的对数谱距离 (Log-Spectral Distance, LSD)。
+    该实现遵循标准的LSD定义: 整体均方根误差。
+
+    Args:
+        reference (np.ndarray): 原始的、干净的参考信号 (一维数组)。
+        estimate (np.ndarray): 模型估计或处理后的信号 (一维数组)。
+        n_fft (int): FFT点数, 决定了频率分辨率。
+        hop_length (int): 帧移, 决定了时间分辨率。
+        win_length (int, optional): 窗长。如果为None, 则默认为n_fft。
+        epsilon (float): 一个非常小的数值, 用于防止对零取对数, 保证数值稳定性。
+
+    Returns:
+        float: 对数谱距离值, 单位为分贝 (dB)。
+    """
+    assert reference.ndim == 1 and estimate.ndim == 1, "输入信号必须是一维数组。"
+    
+    if win_length is None:
+        win_length = n_fft
+
+    reference_stft = librosa.stft(reference, n_fft=n_fft, hop_length=hop_length, win_length=win_length) # (F,T)
+    estimate_stft = librosa.stft(estimate, n_fft=n_fft, hop_length=hop_length, win_length=win_length) # (F,T)
+
+    reference_power_spec = np.abs(reference_stft) ** 2 # (F,T)
+    estimate_power_spec = np.abs(estimate_stft) ** 2 # (F,T)
+
+    reference_log_power_spec = 10 * np.log10(reference_power_spec + EPS)
+    estimate_log_power_spec = 10 * np.log10(estimate_power_spec + EPS)
+
+    squared_error = (reference_log_power_spec - estimate_log_power_spec) ** 2
+    lsd_val = np.sqrt(np.mean(squared_error))
+
+    return lsd_val
+
+def mcd(ref_wav, test_wav, sr=16000):
+    """
+    梅尔倒谱距离 Mel-Cepstral Distance
+    ref_spec: 参考频谱
+    test_spec: 测试频谱
+    """
+    ref_wav, ref_sr = sf.read(ref_wav)
+    test_wav, test_sr = sf.read(test_wav)
+    assert ref_sr == test_sr == sr, "采样率必须为16000Hz"
+    assert len(ref_wav) == len(test_wav), "音频长度必须相同"
+    
+    ref_mfcc = librosa.feature.mfcc(y=ref_wav, sr=sr)
+    test_mfcc = librosa.feature.mfcc(y=test_wav, sr=sr)
+    
+    # 计算 MCD (跳过 0 阶)
+    diff = ref_mfcc[1:] - test_mfcc[1:]
+    mcd = (10.0 / np.log(10)) * np.sqrt(2 * np.mean(np.sum(diff ** 2, axis=0)))
+    return mcd

neverlib/.history/metrics/spec_metric_20250816135653.py

@@ -0,0 +1,65 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-08-16 13:51:57
+Description: 
+'''
+
+import librosa
+import numpy as np
+import soundfile as sf
+from utils import EPS
+
+
+def lsd(ref_wav, test_wav, n_fft=2048, hop_length=512, win_length=None):
+    """
+    计算两个一维音频信号之间的对数谱距离 (Log-Spectral Distance, LSD)。
+    该实现遵循标准的LSD定义: 整体均方根误差。
+
+    Args:
+        ref_wav (np.ndarray): 原始的、干净的参考信号 (一维数组)。
+        test_wav (np.ndarray): 模型估计或处理后的信号 (一维数组)。
+        n_fft (int): FFT点数, 决定了频率分辨率。
+        hop_length (int): 帧移, 决定了时间分辨率。
+        win_length (int, optional): 窗长。如果为None, 则默认为n_fft。
+        epsilon (float): 一个非常小的数值, 用于防止对零取对数, 保证数值稳定性。
+
+    Returns:
+        float: 对数谱距离值, 单位为分贝 (dB)。
+    """
+    assert ref_wav.ndim == 1 and test_wav.ndim == 1, "输入信号必须是一维数组。"
+    
+    if win_length is None:
+        win_length = n_fft
+
+    ref_stft = librosa.stft(ref_wav, n_fft=n_fft, hop_length=hop_length, win_length=win_length) # (F,T)
+    test_stft = librosa.stft(test_wav, n_fft=n_fft, hop_length=hop_length, win_length=win_length) # (F,T)
+
+    ref_power_spec = np.abs(ref_stft) ** 2 # (F,T)
+    test_power_spec = np.abs(test_stft) ** 2 # (F,T)
+
+    ref_log_power_spec = 10 * np.log10(ref_power_spec + EPS)
+    test_log_power_spec = 10 * np.log10(test_power_spec + EPS)
+
+    squared_error = (ref_log_power_spec - test_log_power_spec) ** 2
+    lsd_val = np.sqrt(np.mean(squared_error))
+
+    return lsd_val
+
+def mcd(ref_wav, test_wav, sr=16000):
+    """
+    梅尔倒谱距离 Mel-Cepstral Distance
+    ref_spec: 参考频谱
+    test_spec: 测试频谱
+    """
+    ref_wav, ref_sr = sf.read(ref_wav)
+    test_wav, test_sr = sf.read(test_wav)
+    assert ref_sr == test_sr == sr, "采样率必须为16000Hz"
+    assert len(ref_wav) == len(test_wav), "音频长度必须相同"
+    
+    ref_mfcc = librosa.feature.mfcc(y=ref_wav, sr=sr)
+    test_mfcc = librosa.feature.mfcc(y=test_wav, sr=sr)
+    
+    # 计算 MCD (跳过 0 阶)
+    diff = ref_mfcc[1:] - test_mfcc[1:]
+    mcd = (10.0 / np.log(10)) * np.sqrt(2 * np.mean(np.sum(diff ** 2, axis=0)))
+    return mcd

neverlib/.history/vad/PreProcess_20250805234211.py

@@ -0,0 +1,63 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-02-13 20:06:07
+LastEditTime: 2025-03-17 16:06:11
+FilePath: \neverlib\vad\PreProcess.py
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/14
+"""
+通过一些预处理方法, 来提高VAD的准确率
+"""
+import numpy as np
+import noisereduce as nr
+
+
+def pre_emphasis(audio_data, alpha=0.97):
+    # y(n)=x(n)−α⋅x(n−1)
+    emphasized_audio = np.append(audio_data[0], audio_data[1:] - alpha * audio_data[:-1])
+    return emphasized_audio
+
+
+def NS(wav, sr=16000, stationary=True, prop_decrease=1.):
+    """ 传统降噪 Doc: https://pypi.org/project/noisereduce/
+    :param wav: (xxx,) or (channels, xxx)
+    :param sr: 采样率
+    :param stationary: 平稳降噪还是非平稳降噪
+    :param prop_decrease: 0~1, 降噪噪声百分比
+    :return:
+    """
+    if stationary:
+        # 平稳噪声抑制 stationary=True
+        reduced_noise = nr.reduce_noise(y=wav, sr=sr, stationary=True,
+                                        prop_decrease=prop_decrease,  # 降噪噪声的比例
+                                        )
+    else:
+        # 非平稳噪声抑制 stationary=False
+        reduced_noise = nr.reduce_noise(y=wav, sr=sr, stationary=False,
+                                        prop_decrease=prop_decrease,
+                                        )
+    return reduced_noise
+
+
+def NS_test():
+    import soundfile as sf
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    wav_NS = NS(wav, sr=sr, stationary=True, prop_decrease=0.6)
+    sf.write("../../wav_data/000_short_NS.wav", wav_NS, samplerate=sr)
+
+    # 绘制降噪后的频谱图
+    import matplotlib.pyplot as plt
+    plt.subplot(211)
+    plt.specgram(wav, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.subplot(212)
+    plt.specgram(wav_NS, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.show()
+
+
+if __name__ == "__main__":
+    NS_test()

neverlib/.history/vad/PreProcess_20250809232455.py

@@ -0,0 +1,63 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-02-13 20:06:07
+LastEditTime: 2025-03-17 16:06:11
+FilePath: \neverlib\vad\PreProcess.py
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/14
+"""
+通过一些预处理方法, 来提高VAD的准确率
+"""
+import numpy as np
+import noisereduce as nr
+
+
+def pre_emphasis(audio_data, alpha=0.97):
+    # y(n)=x(n)−α⋅x(n−1)
+    emphasized_audio = np.append(audio_data[0], audio_data[1:] - alpha * audio_data[:-1])
+    return emphasized_audio
+
+
+def NS(wav, sr=16000, stationary=True, prop_decrease=1.):
+    """ 传统降噪 Doc: https://pypi.org/project/noisereduce/
+    :param wav: (xxx,) or (channels, xxx)
+    :param sr: 采样率
+    :param stationary: 平稳降噪还是非平稳降噪
+    :param prop_decrease: 0~1, 降噪噪声百分比
+    :return:
+    """
+    if stationary:
+        # 平稳噪声抑制 stationary=True
+        reduced_noise = nr.reduce_noise(y=wav, sr=sr, stationary=True,
+                                        prop_decrease=prop_decrease,  # 降噪噪声的比例
+                                        )
+    else:
+        # 非平稳噪声抑制 stationary=False
+        reduced_noise = nr.reduce_noise(y=wav, sr=sr, stationary=False,
+                                        prop_decrease=prop_decrease,
+                                        )
+    return reduced_noise
+
+
+def NS_test():
+    import soundfile as sf
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    wav_NS = NS(wav, sr=sr, stationary=True, prop_decrease=0.6)
+    sf.write("../../wav_data/000_short_NS.wav", wav_NS, samplerate=sr)
+
+    # 绘制降噪后的频谱图
+    import matplotlib.pyplot as plt
+    plt.subplot(211)
+    plt.specgram(wav, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.subplot(212)
+    plt.specgram(wav_NS, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.show()
+
+
+if __name__ == "__main__":
+    NS_test()

neverlib/.history/vad/PreProcess_20250816020725.py

@@ -0,0 +1,66 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-02-13 20:06:07
+LastEditTime: 2025-08-16 02:07:24
+FilePath: \neverlib\vad\PreProcess.py
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/14
+"""
+通过一些预处理方法, 来提高VAD的准确率
+"""
+import numpy as np
+import noisereduce as nr
+
+
+def pre_emphasis(audio_data, alpha=0.97):
+    """
+    预加重
+    """
+    # y(n)=x(n)−α⋅x(n−1)
+    emphasized_audio = np.append(audio_data[0], audio_data[1:] - alpha * audio_data[:-1])
+    return emphasized_audio
+
+
+def NS(wav, sr=16000, stationary=True, prop_decrease=1.):
+    """ 传统降噪 Doc: https://pypi.org/project/noisereduce/
+    :param wav: (xxx,) or (channels, xxx)
+    :param sr: 采样率
+    :param stationary: 平稳降噪还是非平稳降噪
+    :param prop_decrease: 0~1, 降噪噪声百分比
+    :return:
+    """
+    if stationary:
+        # 平稳噪声抑制 stationary=True
+        reduced_noise = nr.reduce_noise(y=wav, sr=sr, stationary=True,
+                                        prop_decrease=prop_decrease,  # 降噪噪声的比例
+                                        )
+    else:
+        # 非平稳噪声抑制 stationary=False
+        reduced_noise = nr.reduce_noise(y=wav, sr=sr, stationary=False,
+                                        prop_decrease=prop_decrease,
+                                        )
+    return reduced_noise
+
+
+def NS_test():
+    import soundfile as sf
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    wav_NS = NS(wav, sr=sr, stationary=True, prop_decrease=0.6)
+    sf.write("../../wav_data/000_short_NS.wav", wav_NS, samplerate=sr)
+
+    # 绘制降噪后的频谱图
+    import matplotlib.pyplot as plt
+    plt.subplot(211)
+    plt.specgram(wav, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.subplot(212)
+    plt.specgram(wav_NS, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.show()
+
+
+if __name__ == "__main__":
+    NS_test()

neverlib/.history/vad/VAD_Silero_20250805234211.py

@@ -0,0 +1,50 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/19
+"""
+
+"""
+import torch
+
+
+class Silero_VAD_C():
+    def __init__(self, sr=16000, threshold=0.5, min_speech_duration_ms=10,
+                 min_silence_duration_ms=140, window_size_samples=512, speech_pad_ms=0):
+        self.sr = sr
+        self.threshold = threshold
+        self.min_speech_duration_ms = min_speech_duration_ms  # 语音块的最小持续时间 ms
+        self.min_silence_duration_ms = min_silence_duration_ms  # 语音块之间的最小静音时间 ms
+        self.window_size_samples = window_size_samples  # 512\1024\1536
+        self.speech_pad_ms = speech_pad_ms  # 最后的语音块由两侧的speech_pad_ms填充
+
+        self.model, utils = torch.hub.load(repo_or_dir='snakers4/silero-vad', model='silero_vad', force_reload=False,
+                                           onnx=True)
+        (self.get_speech_timestamps, save_audio, read_audio, VADIterator, collect_chunks) = utils
+
+    def process(self, wav):
+        assert wav.ndim == 1, f"wav shape为{wav.shape}, 期望1D"
+        speech_timestamps = self.get_speech_timestamps(wav, self.model,
+                                                       sampling_rate=self.sr,
+                                                       threshold=self.threshold,
+                                                       min_speech_duration_ms=self.min_speech_duration_ms,
+                                                       min_silence_duration_ms=self.min_silence_duration_ms,
+                                                       window_size_samples=self.window_size_samples,
+                                                       speech_pad_ms=self.speech_pad_ms,
+                                                       )
+        return speech_timestamps
+
+
+if __name__ == "__main__":
+    import soundfile as sf
+    from neverlib.vad.PreProcess import HPFilter, volume_norm
+
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    assert wav_sr == sr, f"音频采样率为{wav_sr}, 期望{sr}"
+    wav = HPFilter(wav, sr=sr, order=6, cutoff=100)
+    wav = volume_norm(wav)
+
+    vad = Silero_VAD_C()
+    vad_array = vad.process(wav)
+    print(vad_array)

neverlib/.history/vad/VAD_Silero_20250809232456.py

@@ -0,0 +1,50 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/19
+"""
+
+"""
+import torch
+
+
+class Silero_VAD_C():
+    def __init__(self, sr=16000, threshold=0.5, min_speech_duration_ms=10,
+                 min_silence_duration_ms=140, window_size_samples=512, speech_pad_ms=0):
+        self.sr = sr
+        self.threshold = threshold
+        self.min_speech_duration_ms = min_speech_duration_ms  # 语音块的最小持续时间 ms
+        self.min_silence_duration_ms = min_silence_duration_ms  # 语音块之间的最小静音时间 ms
+        self.window_size_samples = window_size_samples  # 512\1024\1536
+        self.speech_pad_ms = speech_pad_ms  # 最后的语音块由两侧的speech_pad_ms填充
+
+        self.model, utils = torch.hub.load(repo_or_dir='snakers4/silero-vad', model='silero_vad', force_reload=False,
+                                           onnx=True)
+        (self.get_speech_timestamps, save_audio, read_audio, VADIterator, collect_chunks) = utils
+
+    def process(self, wav):
+        assert wav.ndim == 1, f"wav shape为{wav.shape}, 期望1D"
+        speech_timestamps = self.get_speech_timestamps(wav, self.model,
+                                                       sampling_rate=self.sr,
+                                                       threshold=self.threshold,
+                                                       min_speech_duration_ms=self.min_speech_duration_ms,
+                                                       min_silence_duration_ms=self.min_silence_duration_ms,
+                                                       window_size_samples=self.window_size_samples,
+                                                       speech_pad_ms=self.speech_pad_ms,
+                                                       )
+        return speech_timestamps
+
+
+if __name__ == "__main__":
+    import soundfile as sf
+    from neverlib.vad.PreProcess import HPFilter, volume_norm
+
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    assert wav_sr == sr, f"音频采样率为{wav_sr}, 期望{sr}"
+    wav = HPFilter(wav, sr=sr, order=6, cutoff=100)
+    wav = volume_norm(wav)
+
+    vad = Silero_VAD_C()
+    vad_array = vad.process(wav)
+    print(vad_array)

neverlib/.history/vad/VAD_WebRTC_20250805234211.py

@@ -0,0 +1,61 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/19
+"""
+
+"""
+import numpy as np
+
+
+class WebRTC_VAD_C():
+    def __init__(self, sr=16000, window_len=10, mode=1):
+        """
+        :param window_len: 窗长(ms)
+        :param mode:
+        """
+        import webrtcvad
+        self.sr = sr
+        self.vad = webrtcvad.Vad()
+        self.vad.set_mode(mode)  # 0~3
+        self.window_len = int(window_len / 1000 * sr)
+
+    def process(self, wav):
+        assert wav.ndim == 1, f"wav shape为{wav.shape}, 期望1D"
+        # float32 -> int16
+        wav_int16 = (wav * np.iinfo(np.int16).max).astype(np.int16)
+        wav_int16 = wav_int16[:len(wav_int16) - len(wav_int16) % self.window_len]  # (105120, 1)
+        vad_array = np.zeros_like(wav_int16)
+        for i in range(0, len(wav_int16), self.window_len):
+            vad_flag = self.vad.is_speech(wav_int16[i:i + self.window_len].tobytes(), self.sr)
+            vad_array[i:i + self.window_len] = vad_flag
+
+        return vad_array
+
+
+if __name__ == "__main__":
+    import soundfile as sf
+    import matplotlib.pyplot as plt
+    from neverlib.vad.PreProcess import HPFilter, volume_norm
+
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    assert wav_sr == sr, f"音频采样率为{wav_sr}, 期望{sr}"
+    wav = HPFilter(wav, sr=sr, order=6, cutoff=100)
+    wav = volume_norm(wav)
+
+    vad = WebRTC_VAD_C()
+    vad_array = vad.process(wav)
+
+    plt.figure(figsize=(20, 5))
+    plt.plot(wav)
+    plt.plot(vad_array)
+    plt.grid()
+    plt.show()
+
+    plt.figure(figsize=(20, 5))
+    plt.subplot(2, 1, 1)
+    plt.specgram(wav, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.subplot(2, 1, 2)
+    plt.specgram(vad_array, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.show()

neverlib/.history/vad/VAD_WebRTC_20250809232456.py

@@ -0,0 +1,61 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/19
+"""
+
+"""
+import numpy as np
+
+
+class WebRTC_VAD_C():
+    def __init__(self, sr=16000, window_len=10, mode=1):
+        """
+        :param window_len: 窗长(ms)
+        :param mode:
+        """
+        import webrtcvad
+        self.sr = sr
+        self.vad = webrtcvad.Vad()
+        self.vad.set_mode(mode)  # 0~3
+        self.window_len = int(window_len / 1000 * sr)
+
+    def process(self, wav):
+        assert wav.ndim == 1, f"wav shape为{wav.shape}, 期望1D"
+        # float32 -> int16
+        wav_int16 = (wav * np.iinfo(np.int16).max).astype(np.int16)
+        wav_int16 = wav_int16[:len(wav_int16) - len(wav_int16) % self.window_len]  # (105120, 1)
+        vad_array = np.zeros_like(wav_int16)
+        for i in range(0, len(wav_int16), self.window_len):
+            vad_flag = self.vad.is_speech(wav_int16[i:i + self.window_len].tobytes(), self.sr)
+            vad_array[i:i + self.window_len] = vad_flag
+
+        return vad_array
+
+
+if __name__ == "__main__":
+    import soundfile as sf
+    import matplotlib.pyplot as plt
+    from neverlib.vad.PreProcess import HPFilter, volume_norm
+
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    assert wav_sr == sr, f"音频采样率为{wav_sr}, 期望{sr}"
+    wav = HPFilter(wav, sr=sr, order=6, cutoff=100)
+    wav = volume_norm(wav)
+
+    vad = WebRTC_VAD_C()
+    vad_array = vad.process(wav)
+
+    plt.figure(figsize=(20, 5))
+    plt.plot(wav)
+    plt.plot(vad_array)
+    plt.grid()
+    plt.show()
+
+    plt.figure(figsize=(20, 5))
+    plt.subplot(2, 1, 1)
+    plt.specgram(wav, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.subplot(2, 1, 2)
+    plt.specgram(vad_array, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.show()

neverlib/.history/vad/VAD_funasr_20250805234211.py

@@ -0,0 +1,54 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/19
+"""
+
+"""
+import numpy as np
+
+
+class FunASR_VAD_C():
+    def __init__(self, sr=16000):
+        from funasr import AutoModel
+        self.sr = sr
+        self.model = AutoModel(model="fsmn-vad", model_revision="v2.0.4")
+
+    def process(self, wav):
+        assert wav.ndim == 1, f"wav shape为{wav.shape}, 期望1D"
+        res_list = self.model.generate(input=wav)
+        vad_array = np.zeros_like(wav)
+        for res in res_list:
+            for value_item in res["value"]:
+                beg, end = value_item
+                vad_array[int(beg * self.sr / 1000):int(end * self.sr / 1000)] = 1
+
+        return vad_array
+
+
+if __name__ == "__main__":
+    import soundfile as sf
+    import matplotlib.pyplot as plt
+    from neverlib.vad.PreProcess import HPFilter, volume_norm
+
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    assert wav_sr == sr, f"音频采样率为{wav_sr}, 期望{sr}"
+    wav = HPFilter(wav, sr=sr, order=6, cutoff=100)
+    wav = volume_norm(wav)
+
+    vad = FunASR_VAD_C()
+    vad_array = vad.process(wav)
+
+    plt.figure(figsize=(20, 5))
+    plt.plot(wav)
+    plt.plot(vad_array)
+    plt.grid()
+    plt.show()
+
+    plt.figure(figsize=(20, 5))
+    plt.subplot(2, 1, 1)
+    plt.specgram(wav, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.subplot(2, 1, 2)
+    plt.specgram(vad_array, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.show()

neverlib/.history/vad/VAD_funasr_20250809232456.py

@@ -0,0 +1,54 @@
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2024/9/19
+"""
+
+"""
+import numpy as np
+
+
+class FunASR_VAD_C():
+    def __init__(self, sr=16000):
+        from funasr import AutoModel
+        self.sr = sr
+        self.model = AutoModel(model="fsmn-vad", model_revision="v2.0.4")
+
+    def process(self, wav):
+        assert wav.ndim == 1, f"wav shape为{wav.shape}, 期望1D"
+        res_list = self.model.generate(input=wav)
+        vad_array = np.zeros_like(wav)
+        for res in res_list:
+            for value_item in res["value"]:
+                beg, end = value_item
+                vad_array[int(beg * self.sr / 1000):int(end * self.sr / 1000)] = 1
+
+        return vad_array
+
+
+if __name__ == "__main__":
+    import soundfile as sf
+    import matplotlib.pyplot as plt
+    from neverlib.vad.PreProcess import HPFilter, volume_norm
+
+    sr = 16000
+    wav_path = "../../data/vad_example.wav"
+    wav, wav_sr = sf.read(wav_path, always_2d=False, dtype="float32")
+    assert wav_sr == sr, f"音频采样率为{wav_sr}, 期望{sr}"
+    wav = HPFilter(wav, sr=sr, order=6, cutoff=100)
+    wav = volume_norm(wav)
+
+    vad = FunASR_VAD_C()
+    vad_array = vad.process(wav)
+
+    plt.figure(figsize=(20, 5))
+    plt.plot(wav)
+    plt.plot(vad_array)
+    plt.grid()
+    plt.show()
+
+    plt.figure(figsize=(20, 5))
+    plt.subplot(2, 1, 1)
+    plt.specgram(wav, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.subplot(2, 1, 2)
+    plt.specgram(vad_array, Fs=sr, scale_by_freq=True, sides='default', cmap="jet")
+    plt.show()