探究音频丢字位置和丢字时间对pesq分数的影响

news2025/1/2 23:07:50

丢字的本质

丢字的本质是在一段音频中一小段数据变为0

丢字对主观感受的影响

1. 丢字位置

丢字的位置对感知效果有很大影响。如果丢字发生在音频信号的静音部分或低能量部分，感知可能不明显；而如果丢字发生在高能量部分或关键音素上，感知会非常明显。

2. 丢字持续时间

虽然10ms的丢字时间相对较短，但如果丢字发生在关键音素或瞬态（如爆破音、元音等）上，感知会更加明显。

3. 音频内容

不同类型的音频内容对丢字的敏感度不同。例如，语音信号中的丢字可能比音乐信号中的丢字更容易被感知，因为语音信号中有更多的瞬态和关键音素。

4. 人耳的感知能力

人耳对不同频率和时间的变化有不同的敏感度。某些频率范围内的丢字可能更容易被感知，而其他频率范围内的丢字可能不明显。

丢字位置和丢字持续时间的影响

判断丢字的位置在高能量和低能量位置以及丢字时间对pesq分数的影响

选取一段音频，随机在其高能量和低能量位置丢字，丢字时间分别设置为

[0.001, 0.002, 0.003, 0.004, 0.005, 0.006, 0.007, 0.008, 0.009, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.1]

单位为s，生成所有丢字的音频，再对丢字音频进行pesq评分，画成折线图输出

脚本代码如下：

import numpy as np
from scipy.io import wavfile
from pesq import pesq
from pesq import PesqError
import librosa
import matplotlib.pyplot as plt

def create_single_drop_audio(data, drop_start, drop_duration, sample_rate):
    """在音频信号中指定位置，并将该位置的一小段音频数据设置为零"""
    num_samples = len(data)
    drop_samples = int(drop_duration * sample_rate)
    drop_end = drop_start + drop_samples
    print(drop_start,drop_duration)
    # 创建丢字音频
    dropped_data = np.copy(data)
    dropped_data[drop_start:drop_end] = 0
    
    return dropped_data

# 读取原始音频文件并转换采样率
original_file = 'audio_file.wav'
target_sample_rate = 16000  # 选择8000或16000

# 使用librosa加载音频文件并转换采样率
original_data, original_sample_rate = librosa.load(original_file, sr=target_sample_rate)

# 计算音频信号的能量分布
energy = np.abs(original_data)**2
window_size = int(0.01 * original_sample_rate)  # 10ms窗口
energy = np.convolve(energy, np.ones(window_size), 'same')

# 随机选择一个低能量位置进行丢字
low_energy_indices = np.where(energy < np.percentile(energy, 20))[0]  # 选择能量最低的20%
high_energy_indices = np.where(energy > np.percentile(energy, 80))[0]  # 选择能量最高的20%

# 定义不同的drop_duration值
drop_durations = [0.001, 0.002, 0.003, 0.004, 0.005, 0.006, 0.007, 0.008, 0.009, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.1]

# 存储PESQ分数
pesq_scores_low_energy = []
pesq_scores_high_energy = []

# 计算原始音频的PESQ分数（与自身比较）
try:
    original_pesq_score = pesq(original_sample_rate, original_data, original_data, 'wb')
    print(f'Original Audio PESQ Score: {original_pesq_score:.2f}')
except PesqError as e:
    print(f'Error calculating PESQ for original audio: {e}')
    original_pesq_score = None

# 对低能量部分进行丢字
drop_start = np.random.choice(low_energy_indices)
for drop_duration in drop_durations:

    dropped_data = create_single_drop_audio(original_data, drop_start, drop_duration, sample_rate=original_sample_rate)
    
    # 保存丢字音频
    output_file = f'low_energy_dropped_audio_{int(drop_duration*1000)}ms.wav'
    wavfile.write(output_file, original_sample_rate, (dropped_data * 32767).astype(np.int16))
    
    try:
        pesq_score = pesq(original_sample_rate, original_data, dropped_data, 'wb')
        pesq_scores_low_energy.append(pesq_score)
        print(f'Low Energy - Drop Duration: {drop_duration:.3f}s, PESQ Score: {pesq_score:.2f}')
    except PesqError as e:
        print(f'Error calculating PESQ for drop_duration {drop_duration} in low energy: {e}')
        pesq_scores_low_energy.append(None)

# 对高能量部分进行丢字
drop_start = np.random.choice(high_energy_indices)
for drop_duration in drop_durations:

    dropped_data = create_single_drop_audio(original_data, drop_start, drop_duration, sample_rate=original_sample_rate)
    
    # 保存丢字音频
    output_file = f'high_energy_dropped_audio_{int(drop_duration*1000)}ms.wav'
    wavfile.write(output_file, original_sample_rate, (dropped_data * 32767).astype(np.int16))
    
    try:
        pesq_score = pesq(original_sample_rate, original_data, dropped_data, 'wb')
        pesq_scores_high_energy.append(pesq_score)
        print(f'High Energy - Drop Duration: {drop_duration:.3f}s, PESQ Score: {pesq_score:.2f}')
    except PesqError as e:
        print(f'Error calculating PESQ for drop_duration {drop_duration} in high energy: {e}')
        pesq_scores_high_energy.append(None)

# 绘制折线图
plt.figure(figsize=(12, 8))
plt.plot([0] + drop_durations, [original_pesq_score] + pesq_scores_low_energy, marker='o', linestyle='-', color='b', label='Low Energy PESQ Score')
plt.plot([0] + drop_durations, [original_pesq_score] + pesq_scores_high_energy, marker='o', linestyle='-', color='r', label='High Energy PESQ Score')
plt.xlabel('Drop Duration (s)')
plt.ylabel('PESQ Score')
plt.title('PESQ Score vs Drop Duration (Low Energy vs High Energy)')
plt.grid(True)
plt.legend()
plt.show()

运行三次，随机选择不同的高能量和低能量部分，生成的折线图