neverlib 0.2.4__py3-none-any.whl → 0.2.6__py3-none-any.whl

neverlib/.history/audio_aug/audio_aug_20250827172347.py

@@ -0,0 +1,471 @@
+# -*- 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
+from neverlib.filter import HPFilter
+
+
+def limiter(wav, threshold=0.999):
+    """
+    简单限幅器, threshold=0.999 ~ -0.1dBFS
+    超过阈值的样本被压缩到阈值限制
+    """
+    peak = np.max(np.abs(wav))
+    if peak > threshold:
+        scalar = threshold / (peak + EPS)
+        wav = wav * scalar
+    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
+        # 防止削波
+        if np.max(np.abs(wav)) > 1:
+            scale_factor = 1 / np.max(np.abs(wav))
+            wav *= scale_factor
+    return wav
+
+
+def snr_aug_changeNoise(clean, noise, target_snr, hpf=False, sr=16000, order=4, cutoff=100):
+    """ 
+    保持语音不变, 改变噪声的幅度
+    HP: 高通滤波, 如果你的数据工频干扰较高, 建议设置为True, 否则snr会不准
+    snr = 10 * log10(signal_power / k*noise_power) 
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_tmp, noise_tmp = clean.copy(), noise.copy()
+    if hpf:
+        clean_tmp = HPFilter(clean_tmp, sr=sr, order=order, cutoff=cutoff)
+        noise_tmp = HPFilter(noise_tmp, sr=sr, order=order, cutoff=cutoff)
+    # 纯净语音功率, 噪声功率
+    clean_power, noise_power = np.mean(clean_tmp ** 2), np.mean(noise_tmp ** 2)
+    noise_scale = np.sqrt(clean_power / (noise_power * 10 ** (target_snr / 10) + EPS))
+    noisy = clean + noise_scale * noise
+    # 防止削波
+    if np.max(np.abs(noisy)) > 1:
+        scale_factor = 1 / np.max(np.abs(noisy))
+        noisy *= scale_factor
+        clean *= scale_factor
+    return noisy, clean
+
+
+def snr_aug_changeClean(clean, noise, target_snr, clip_check=True, hpf=False, sr=16000, order=4, cutoff=100):
+    """ 
+    保持噪声不变，改变纯净语音的幅度以达到目标信噪比
+    snr = 10 * log10(k*signal_power/ noise_power)
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_tmp, noise_tmp = clean.copy(), noise.copy()
+    if hpf:
+        clean_tmp = HPFilter(clean_tmp, sr=sr, order=order, cutoff=cutoff)
+        noise_tmp = HPFilter(noise_tmp, sr=sr, order=order, cutoff=cutoff)
+    # 纯净语音功率, 噪声功率
+    clean_power, noise_power = np.mean(clean_tmp ** 2), np.mean(noise_tmp ** 2)
+    # 计算纯净信号需要的幅度因子
+    clean_scale = np.sqrt(noise_power * 10 ** (target_snr / 10) / (clean_power + EPS))
+    noisy = clean * clean_scale + noise
+    # 防止削波
+    if clip_check:
+        if np.max(np.abs(noisy)) > 1:
+            scale_factor = 1 / np.max(np.abs(noisy))
+            noisy *= scale_factor
+            clean *= (scale_factor * clean_scale)
+    return noisy, clean
+
+
+def snr_diff_changeNoise(clean, noise, target_snr, hpf=False, sr=16000, order=4, cutoff=100):
+    """ 
+    保持语音不变, 改变噪声的幅度, 和snr_aug_changeNoise作用等效
+    snr = 10 * log10(signal_power / k*noise_power) 
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_tmp, noise_tmp = clean.copy(), noise.copy()
+    if hpf:
+        clean_tmp = HPFilter(clean_tmp, sr=sr, order=order, cutoff=cutoff)
+        noise_tmp = HPFilter(noise_tmp, sr=sr, order=order, cutoff=cutoff)
+    # 纯净语音功率, 噪声功率
+    clean_power, noise_power = np.mean(clean_tmp ** 2), np.mean(noise_tmp ** 2)
+    source_snr = 10 * np.log10(clean_power / (noise_power + EPS) + EPS)
+    noise_dB = source_snr - target_snr     # 噪声还差多少dB
+    noise_gain = 10 ** (noise_dB / 20)
+    noisy = clean + noise_gain * noise
+    # 防止削波
+    if np.max(np.abs(noisy)) > 1:
+        scale_factor = 1 / np.max(np.abs(noisy))
+        noisy *= scale_factor
+        clean *= scale_factor
+    return noisy, clean
+
+
+def snr_diff_changeClean(clean, noise, target_snr, clip_check=True, hpf=False, sr=16000, order=4, cutoff=100):
+    """ 
+    保持噪声不变, 改变纯净语音的幅度, 和snr_aug_changeClean作用等效
+    snr = 10 * log10(signal_power / k*noise_power) 
+    """
+    assert clean.shape == noise.shape, "clean and noise must have the same shape"
+    clean_tmp, noise_tmp = clean.copy(), noise.copy()
+    if hpf:
+        clean_tmp = HPFilter(clean_tmp, sr=sr, order=order, cutoff=cutoff)
+        noise_tmp = HPFilter(noise_tmp, sr=sr, order=order, cutoff=cutoff)
+    # 纯净语音功率, 噪声功率
+    clean_power, noise_power = np.mean(clean_tmp ** 2), np.mean(noise_tmp ** 2)
+    source_snr = 10 * np.log10(clean_power / (noise_power + EPS) + EPS)
+    clean_dB = target_snr - source_snr     # 纯净语音还差多少dB
+    clean_gain = 10 ** (clean_dB / 20)
+    noisy = clean_gain * clean + noise
+    # 防止削波
+    if clip_check:
+        if np.max(np.abs(noisy)) > 1:
+            scale_factor = 1 / np.max(np.abs(noisy))
+            noisy *= scale_factor
+            clean *= (scale_factor * clean_gain)
+    return noisy, clean
+
+
+def snr_aug_Interpolation(clean, noise, target_snr, hpf=False, sr=16000, order=4, cutoff=100):
+    """
+    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."
+    clean_tmp, noise_tmp = clean.copy(), noise.copy()
+    if hpf:
+        clean_tmp = HPFilter(clean_tmp, sr=sr, order=order, cutoff=cutoff)
+        noise_tmp = HPFilter(noise_tmp, sr=sr, order=order, cutoff=cutoff)
+    noisy = noise.copy()
+    index = random.randint(0, noise_len - clean_len)
+    noise_tmp = noise_tmp[index:index + clean_len, :]
+    # 这里必须把clip_check设置为False, 否则外面的noise和里面的不一致
+    noisy_tmp, clean_tmp = snr_aug_changeClean(clean_tmp, noise_tmp, target_snr, clip_check=False, hpf=False)
+    noisy[index:index + clean_len, :] = noisy_tmp
+    # 防止削波
+    if np.max(np.abs(noisy)) > 1:
+        scale_factor = 1 / np.max(np.abs(noisy))
+        noisy *= scale_factor
+        clean *= scale_factor
+    return noisy, np.pad(clean, ((index, noise_len - index - clean_len), (0, 0)))
+
+
+def get_snr_use_vad(wav, vad, sr=16000):
+    # 通过vad获得语音原始的snr
+    wav = HPFilter(wav, sr=sr, order=6, cutoff=100)
+    vadstart, vadend = vad["start"], vad["end"]
+    noise = np.concatenate([wav[:vadstart], wav[vadend:]], axis=0)
+    speech_segment = wav[vadstart:vadend]
+
+    # 计算信噪比
+    # 统计语音段的均方功率谱
+    P_speech_noise = np.mean(speech_segment ** 2)  # 语音+噪声功率
+    # P_speech_noise = np.mean(wav ** 2)  # 如果用全局的, 会存在噪声功率过大的问题, 导致snr过低
+    P_noise = np.mean(noise ** 2)  # 纯噪声功率
+
+    # 计算净语音功率(确保非负)
+    P_speech = max(P_speech_noise - P_noise, 1e-8)
+    if P_noise < 1e-8:
+        P_noise = 1e-8
+
+    snr = 10 * np.log10(P_speech / P_noise)  # 计算 SNR
+    # snr保留小数点后一位
+    return round(snr, 1)
+
+
+def snr_aug_vad_Interpolation(clean, noise, target_snr, vad, hpf=False, sr=16000, order=4, cutoff=100):
+    """
+    在已知clean_len<=noise_len的情况下, 将clean插入到noise中的snr aug方法,
+    使用VAD信息, 精确地找到语音位置
+    Args:
+        clean: 语音
+        noise: 噪声
+        vad: {"start": xxx, "end": xxx}
+    """
+    clean_len, noise_len = clean.shape[0], noise.shape[0]
+    assert clean_len <= noise_len, f"clean_len must be less than noise_len."
+    clean_tmp, noise_tmp = clean.copy(), noise.copy()
+    noisy = noise.copy()
+    index = random.randint(0, noise_len - clean_len)
+    noise = noise[index:index + clean_len, :]  # 现在语音和噪声长度一致
+    if hpf:
+        clean_tmp = HPFilter(clean_tmp, sr=sr, order=order, cutoff=cutoff)
+        noise_tmp = HPFilter(noise_tmp, sr=sr, order=order, cutoff=cutoff)
+    # 只根据语音段求SNR
+    clean_vad = clean_tmp[vad["start"]:vad["end"]]
+    noise_tmp = noise_tmp[vad["start"]:vad["end"]]
+    power_clean, power_noise = np.mean(clean_vad ** 2), np.mean(noise_tmp ** 2)
+    snr_in = 10 * np.log10(power_clean / (power_noise + EPS) + EPS)
+    clean_dB = target_snr - snr_in  # 语音还差多少dB
+    # noise_dB = snr_in - target_snr    # 噪声还差多少dB
+    clean_gain = 10 ** (clean_dB / 20)
+    noisy_tmp = clean_gain * clean + noise
+
+    noisy[index:index + clean_len, :] = noisy_tmp
+    # 防止削波
+    if np.max(np.abs(noisy)) > 1:
+        scale_factor = 1 / np.max(np.abs(noisy))
+        noisy *= scale_factor
+        clean *= (scale_factor * clean_gain)
+    return noisy, np.pad(clean, ((index, noise_len - index - clean_len), (0, 0)))
+
+
+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
+
+
+# ----------------------------------------------------------------
+# 音量增强
+# ----------------------------------------------------------------
+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 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))
+            scalar = target_level / (ipt_max + EPS)
+        else:
+            raise ValueError("method must be 'dbrms' or 'linmax'")
+        wav *= scalar
+    return wav
+
+
+def volume_aug_dbrms(wav, target_level, hpf=False, sr=16000, order=4, cutoff=100):
+    """
+    音量增强, 使用dbrms方法
+    Args:
+        wav: 音频
+        target_level: 目标音量, 单位dB
+        hpf: 是否高通滤波
+        sr: 采样率
+        order: 滤波器阶数
+        cutoff: 截止频率
+    """
+    wav_tmp = wav.copy()
+    if hpf:
+        wav_tmp = HPFilter(wav_tmp, sr=sr, order=order, cutoff=cutoff)
+    wav_rms = (wav_tmp ** 2).mean() ** 0.5
+    scalar = 10 ** (target_level / 20) / (np.max(wav_rms) + EPS)
+    wav_opt = wav_tmp * scalar
+    wav_opt = limiter(wav_opt)
+    return wav_opt
+
+
+def volume_aug_linmax(wav, target_level, hpf=False, sr=16000, order=4, cutoff=100):
+    """
+    音量增强, 使用linmax方法
+    Args:
+        wav: 音频
+        target_level: 目标音量, 单位dB
+        hpf: 是否高通滤波
+        sr: 采样率
+        order: 滤波器阶数
+        cutoff: 截止频率
+    """
+    assert target_level > 0 and target_level < 1, "target_level must be between 0 and 1"
+    wav_tmp = wav.copy()
+    if hpf:
+        wav_tmp = HPFilter(wav_tmp, sr=sr, order=order, cutoff=cutoff)
+    wav_max = np.max(np.abs(wav_tmp))
+    scalar = target_level / (wav_max + EPS)
+    wav *= scalar
+    wav = limiter(wav)
+
+    return wav
+
+
+import pyloudnorm as pyln
+
+def volume_aug_lufs(wav, target_lufs, hpf=False, sr=16000, order=4, cutoff=100):
+    """
+    音量增强, 使用lufs方法, 
+    LUFS是“感知响度” → 跟人耳听感对齐，而且符合国际响度标准。
+    
+    LUFS 使用 感知加权（K-weighting）
+    - 高频增强（模拟人耳在 3~6kHz 的敏感）
+    - 低频衰减（降低 <100Hz 对响度的影响）。
+    使用 短时块（400ms）能量 + 响度门限（-70 LUFS） 过滤极静音段。
+    
+    Args:
+        wav: 音频
+        target_lufs: 目标音量, 单位lufs
+        hpf: 是否高通滤波
+        sr: 采样率
+        order: 滤波器阶数
+        cutoff: 截止频率
+        
+    补充信息: 
+    ## 推荐的 target_lufs 值（行业参考）
+    平台	推荐目标 LUFS
+    YouTube / Spotify	-14
+    Apple Music	-16
+    广播 / TV	-23
+    游戏音频	-16 ~ -18
+    有声书	-18 ~ -20
+    """
+    wav_tmp = wav.copy()
+    if hpf:
+        wav_tmp = HPFilter(wav_tmp, sr=sr, order=4, cutoff=1000)
+
+    # Step2: 创建 LUFS 测量器（ITU-R BS.1770）
+    meter = pyln.Meter(sr, block_size=0.400)  # block_size=400ms
+
+    # Step3: 测量当前 LUFS
+    loudness = meter.integrated_loudness(wav_tmp)   
+
+    # Step4: 计算增益并应用
+    loudness_diff = target_lufs - loudness
+    scalar = 10 ** (loudness_diff / 20.0)
+    wav = wav * scalar
+
+    wav = limiter(wav, threshold=0.999) # Step5: 限幅
+    return wav
+
+def measure_loudness(wav, sr):
+    """
+    测量音频的 Peak / RMS / LUFS，以及峰均比（Crest Factor）
+
+    参数:
+        wav: np.ndarray, 音频波形（范围 [-1, 1]）
+        sr: int, 采样率
+
+    返回:
+        dict:
+            - peak_dbfs: 峰值(dBFS)
+            - rms_dbfs: 均方根电平(dBFS)
+            - lufs: 感知响度（LUFS，ITU-R BS.1770-4标准）
+            - crest_factor_db: 峰均比(dB)，峰值与RMS的差值
+    """
+    EPS = 1e-9
+
+    # 1. Peak
+    peak_linear = np.max(np.abs(wav))
+    peak_dbfs = 20 * np.log10(peak_linear + EPS)
+
+    # 2. RMS
+    rms_val = np.sqrt(np.mean(wav ** 2))
+    rms_dbfs = 20 * np.log10(rms_val + EPS)
+
+    # 3. LUFS
+    meter = pyln.Meter(sr, block_size=0.400)  # 400ms 块
+    loudness_lufs = meter.integrated_loudness(wav)
+
+    # 4. Crest Factor (峰均比)
+    crest_factor_db = peak_dbfs - rms_dbfs
+
+    return {
+        "peak_dbfs": peak_dbfs,
+        "rms_dbfs": rms_dbfs,
+        "lufs": loudness_lufs,
+        "crest_factor_db": crest_factor_db
+    }
+    
+    
+def volume_convert(value, 
+                   from_unit="linear", to_unit="dBFS", 
+                   crest_factor_db=None, lufs_offset=None):
+    """
+    音量单位转换函数
+
+    参数:
+        value: float
+            输入值（可能是线性幅度、dBFS、LUFS）
+        from_unit: str
+            输入单位: "linear", "dBFS", "RMS_dBFS", "LUFS"
+        to_unit: str
+            输出单位: "linear", "dBFS", "RMS_dBFS", "LUFS"
+        crest_factor_db: float | None
+            峰均比（用于 Peak <-> RMS 的转换）
+        lufs_offset: float | None
+            LUFS 与 RMS 的差值（用于 RMS <-> LUFS 转换）
+            例如, 对人声，LUFS ≈ RMS_dBFS - 1.5
+
+    返回:
+        float
+    """
+    EPS = 1e-9
+
+    # Step 1: 统一转换成线性幅度（以满刻度 1.0 为基准）
+    if from_unit == "linear":
+        lin_val = value
+    elif from_unit in ("dBFS", "Peak_dBFS"):
+        lin_val = 10 ** (value / 20)
+    elif from_unit == "RMS_dBFS":
+        lin_val = 10 ** (value / 20)
+    elif from_unit == "LUFS":
+        if lufs_offset is None:
+            raise ValueError("LUFS ↔ RMS 转换需要提供 lufs_offset")
+        rms_dbfs = value + lufs_offset
+        lin_val = 10 ** (rms_dbfs / 20)
+    else:
+        raise ValueError(f"未知单位：{from_unit}")
+
+    # Step 2: 从线性幅度转换到目标单位
+    if to_unit == "linear":
+        return lin_val
+    elif to_unit in ("dBFS", "Peak_dBFS"):
+        return 20 * np.log10(lin_val + EPS)
+    elif to_unit == "RMS_dBFS":
+        if from_unit in ("dBFS", "Peak_dBFS") and crest_factor_db is not None:
+            # 用Peak -> RMS
+            peak_dbfs = 20 * np.log10(lin_val + EPS)
+            return peak_dbfs - crest_factor_db
+        return 20 * np.log10(lin_val + EPS)
+    elif to_unit == "LUFS":
+        if lufs_offset is None:
+            raise ValueError("RMS ↔ LUFS 转换需要提供 lufs_offset")
+        rms_dbfs = 20 * np.log10(lin_val + EPS)
+        return rms_dbfs - lufs_offset
+    else:
+        raise ValueError(f"未知单位：{to_unit}")