基于深度学习的婴儿啼哭识别项目详解

一、项目背景

婴儿啼哭声是婴儿沟通需求的重要信号，对于父母和护理者而言至关重要。本项目基于PaddleSpeech框架，致力于构建婴儿啼哭识别系统，通过深度学习将啼哭声翻译成成人语言，帮助理解婴儿的需求和状态。

1.1 项目背景

婴儿啼哭声是一种生物报警器，传递婴儿的生理和心理需求。有效地识别啼哭声有助于提高婴儿护理的效率和质量。

1.2 数据说明

项目使用六类人工添加噪声的哭声作为训练数据集，分别代表不同的婴儿需求，如苏醒、换尿布、要抱抱、饥饿、困乏、不舒服。噪声数据来自Noisex-92标准数据库。

二、PaddleSpeech环境准备

安装PaddleSpeech和PaddleAudio，确保环境准备就绪。

!python -m pip install -q -U pip --user
!pip install paddlespeech paddleaudio -U -q

三、数据预处理

3.1 数据解压缩

解压缩训练数据集，获取音频文件。

!unzip -qoa data/data41960/dddd.zip

3.2 查看声音文件

通过可视化展示音频波形，了解样本数据的特征。

from paddleaudio import load
data, sr = load(file='train/awake/awake_0.wav', mono=True, dtype='float32')  
print('wav shape: {}'.format(data.shape))
print('sample rate: {}'.format(sr))
plt.figure()
plt.plot(data)
plt.show()

3.3 音频文件长度处理

统一音频文件长度，确保训练数据格式一致。

# 音频信息查看
import soundfile as sf
import numpy as np
import librosa

data, samplerate = sf.read('hungry_0.wav')
channels = len(data.shape)
length_s = len(data) / float(samplerate)
format_rate = 16000
print(f"channels: {channels}")
print(f"length_s: {length_s}")
print(f"samplerate: {samplerate}")

四、自定义数据集与模型训练

4.1 自定义数据集

创建自定义数据集类，包含六类婴儿需求的音频文件。

class CustomDataset(AudioClassificationDataset):
    # List all the class labels
    label_list = [
        'awake',
        'diaper',
        'hug',
        'hungry',
        'sleepy',
        'uncomfortable'
    ]

    train_data_dir = './train/'

    def __init__(self, **kwargs):
        files, labels = self._get_data()
        super(CustomDataset, self).__init__(
            files=files, labels=labels, feat_type='raw', **kwargs)

    # 返回音频文件、label值
    def _get_data(self):
        '''
        This method offer information of wave files and labels.
        '''
        files = []
        labels = []

        for i in range(len(self.label_list)):
            single_class_path = os.path.join(self.train_data_dir, self.label_list[i])
            for sound in os.listdir(single_class_path):
                if 'wav' in sound:
                    sound = os.path.join(single_class_path, sound)
                    files.append(sound)
                    labels.append(i)
        return files, labels

4.2 模型训练

选取预训练模型作为特征提取器，构建分类模型进行模型训练。

# 选取cnn14作为 backbone，用于提取音频的特征
from paddlespeech.cls.models import cnn14
backbone = cnn14(pretrained=True, extract_embedding=True)

# 构建分类模型
class SoundClassifier(nn.Layer):
    def __init__(self, backbone, num_class, dropout=0.1):
        super().__init__()
        self.backbone = backbone
        self.dropout = nn.Dropout(dropout)
        self.fc = nn.Linear(self.backbone.emb_size, num_class)

    def forward(self, x):
        x = x.unsqueeze(1)
        x = self.backbone(x)
        x = self.dropout(x)
        logits = self.fc(x)
        return logits

model = SoundClassifier(backbone, num_class=len(train_ds.label_list))

4.3 模型训练

定义优化器和损失函数，进行模型训练。

# 定义优化器和 Loss
optimizer = paddle.optimizer.Adam(learning_rate=1e-4, parameters=model.parameters())
criterion = paddle.nn.loss.CrossEntropyLoss()

# 模型训练
epochs = 20
steps_per_epoch = len(train_loader)
log_freq = 10
eval_freq = 10

for epoch in range(1, epochs + 1):
    model.train()

    avg_loss = 0
    num_corrects = 0
    num_samples = 0
    
    for batch_idx, batch in enumerate(train_loader):
        waveforms, labels = batch
        feats = feature_extractor(waveforms)
        feats = paddle.transpose(feats, [0, 2, 1])  
        logits = model(feats)

        loss = criterion(logits, labels)
        loss.backward()
        optimizer.step()
        if isinstance(optimizer._learning_rate, paddle.optimizer.lr.LRScheduler):
            optimizer._learning_rate.step()
        optimizer.clear_grad()

        # 计算损失
        avg_loss += loss.numpy()[0]

        # 计算指标
        preds = paddle.argmax(logits, axis=1)
        num_corrects += (preds == labels).numpy().sum()
        num_samples += feats.shape[0]

        if (batch_idx + 1) % log_freq == 0:
            lr = optimizer.get_lr()
            avg_loss /= log_freq
            avg_acc = num_corrects / num_samples



            print_msg = 'Epoch={}/{}, Step={}/{}'.format(
                epoch, epochs, batch_idx + 1, steps_per_epoch)
            print_msg += ' loss={:.4f}'.format(avg_loss)
            print_msg += ' acc={:.4f}'.format(avg_acc)
            print_msg += ' lr={:.6f}'.format(lr)
            logger.train(print_msg)

            avg_loss = 0
            num_corrects = 0
            num_samples = 0

五、模型测试

通过模型对测试音频进行推理，输出对应的婴儿需求概率。

# 模型测试
top_k = 3
wav_file = 'test/test_0.wav'
n_fft = 1024
win_length = 1024
hop_length = 320
f_min = 50.0
f_max = 16000.0

waveform, sr = load(wav_file, sr=sr)
feature_extractor = LogMelSpectrogram(
    sr=sr, 
    n_fft=n_fft, 
    hop_length=hop_length, 
    win_length=win_length, 
    window='hann', 
    f_min=f_min, 
    f_max=f_max, 
    n_mels=64)
feats = feature_extractor(paddle.to_tensor(paddle.to_tensor(waveform).unsqueeze(0)))
feats = paddle.transpose(feats, [0, 2, 1])

logits = model(feats)
probs = nn.functional.softmax(logits, axis=1).numpy()

sorted_indices = probs[0].argsort()

msg = f'[{wav_file}]\n'
for idx in sorted_indices[-1:-top_k-1:-1]:
    msg += f'{train_ds.label_list[idx]}: {probs[0][idx]:.5f}\n'
print(msg)

六、注意事项

1. 自定义数据集格式参考文档；
2. 统一音频尺寸，确保音频长度和采样频率一致；
3. 可学习PaddleSpeech课程。