python PaddleSpeech implementation for baby cry recognition

# Environment preparation: install paddlespeech and paddleaudio
!python -m pip install -q -U pip --user
!pip install paddlespeech paddleaudio -U -q

!pip list|grep paddle

import warnings
("ignore")
import IPython
import numpy as np
import  as plt
import paddle
%matplotlib inline

# !unzip -qoa data/data41960/

from paddleaudio import load
data, sr = load(file='train/awake/awake_0.wav', mono=True, dtype='float32')  # Single channel, float32 audio sample points
print('wav shape: {}'.format())
print('sample rate: {}'.format(sr))
# Show audio waveforms
()
(data)
()

from paddleaudio import load
data, sr = load(file='train/diaper/diaper_0.wav', mono=True, dtype='float32')  # Single channel, float32 audio sample points
print('wav shape: {}'.format())
print('sample rate: {}'.format(sr))
# Show audio waveforms
()
(data)
()

!paddlespeech cls --input train/awake/awake_0.wav

!paddlespeech help

# Check the audio length
import contextlib
import wave
def get_sound_len(file_path):
    with ((file_path, 'r')) as f:
        frames = ()
        rate = ()
        wav_length = frames / float(rate)
    return wav_length

# Compile wav files
import glob
sound_files=('train/*/*.wav')
print(sound_files[0])
print(len(sound_files))

# Counting the longest and shortest audio
sounds_len=[]
for sound in sound_files:
    sounds_len.append(get_sound_len(sound))
print("Maximum length of audio:",max(sounds_len),"Seconds.")
print("Minimum audio length:",min(sounds_len),"Seconds.")

!cp train/hungry/hungry_0.wav ~/

!pip install pydub -q

# Audio information view
import math
import soundfile as sf
import numpy as np
import librosa
data, samplerate = ('hungry_0.wav')
channels = len()
length_s = len(data)/float(samplerate)
format_rate=16000
print(f"channels: {channels}")
print(f"length_s: {length_s}")
print(f"samplerate: {samplerate}")

# Harmonized to 34s
from pydub import AudioSegment
audio = AudioSegment.from_wav('hungry_0.wav')
print(str(audio.duration_seconds))
i = 1
padded = audio
while padded.duration_seconds * 1000 < 34000:
    padded = audio * i
    i = i + 1
padded[0:34000].set_frame_rate(16000).export('', format='wav')

import math
import soundfile as sf
import numpy as np
import librosa
data, samplerate = ('')
channels = len()
length_s = len(data)/float(samplerate)
format_rate=16000
print(f"channels: {channels}")
print(f"length_s: {length_s}")
print(f"samplerate: {samplerate}")

# Define a function that repeats the padding if the maximum length is not reached, ultimately intercepting from more than 34s of audio
from pydub import AudioSegment
def convert_sound_len(filename):
    audio = AudioSegment.from_wav(filename)
    i = 1
    padded = audio*i
    while padded.duration_seconds * 1000 < 34000:
        i = i + 1
        padded = audio * i
    padded[0:34000].set_frame_rate(16000).export(filename, format='wav')

# Harmonize all audio to fixed length
for sound in sound_files:
    convert_sound_len(sound)

import os
from  import AudioClassificationDataset
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)
    # Return audio file, label value
    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=(self.train_data_dir, self.label_list[i])            
            for sound in (single_class_path):
                # print(sound)
                if 'wav' in sound:
                    sound=(single_class_path, sound)
                    (sound)
                    (i)
        return files, labels

# Define dataloader
import paddle
from  import LogMelSpectrogram
# Feature config should be align with pretrained model
sample_rate = 16000
feat_conf = {
  'sr': sample_rate,
  'n_fft': 1024,
  'hop_length': 320,
  'window': 'hann',
  'win_length': 1024,
  'f_min': 50.0,
  'f_max': 14000.0,
  'n_mels': 64,
}
train_ds = CustomDataset(sample_rate=sample_rate)
feature_extractor = LogMelSpectrogram(**feat_conf)
train_sampler = (
    train_ds, batch_size=64, shuffle=True, drop_last=False)
train_loader = (
    train_ds,
    batch_sampler=train_sampler,
    return_list=True,
    use_buffer_reader=True)

from  import cnn14
backbone = cnn14(pretrained=True, extract_embedding=True)

import  as nn
class SoundClassifier():
    def __init__(self, backbone, num_class, dropout=0.1):
        super().__init__()
         = backbone
         = (dropout)
         = (.emb_size, num_class)
    def forward(self, x):
        x = (1)
        x = (x)
        x = (x)
        logits = (x)
        return logits
model = SoundClassifier(backbone, num_class=len(train_ds.label_list))

# Define the Optimizer and Loss
optimizer = (learning_rate=1e-4, parameters=())
criterion = ()

from  import logger
epochs = 20
steps_per_epoch = len(train_loader)
log_freq = 10
eval_freq = 10
for epoch in range(1, epochs + 1):
    ()
    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 = (feats, [0, 2, 1])  # [B, N, T] -> [B, T, N]
        logits = model(feats)
        loss = criterion(logits, labels)
        ()
        ()
        if isinstance(optimizer._learning_rate,
                      ):
            optimizer._learning_rate.step()
        optimizer.clear_grad()
        # Calculate loss
        avg_loss += ()[0]
        # Calculate metrics
        preds = (logits, axis=1)
        num_corrects += (preds == labels).numpy().sum()
        num_samples += [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)
            (print_msg)
            avg_loss = 0
            num_corrects = 0
            num_samples = 0

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 = (feats, [0, 2, 1])  # [B, N, T] -> [B, T, N]
logits = model(feats)
probs = (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)