폴더 안의 파일을 mel-spectrogram으로 바꾸자

우리가 해야 될 일은 다음과 같다.

1. librosa에 load 하고 전체 음원을 구간별로 나누는 클래스

2. 만들 폴더의 경로와 폴더에서 wav파일을 가져오는 클래스

3. 단일 wav 음원을 mel-spectrogram으로 바꾸는 클래스

4. 1, 2, 3을 모두 합쳐 폴더 안의 wav 파일을 mel-spectrogram으로 바꾸는 클래스

모든 클래스는 문서화를 해서 자세한 설명은 문서를 참고하자

1.librosa에 load 하고 전체 음원을 구간별로 나누는 클래스

import librosa


class AudioProcessor:
    """
    A class used to process audio data, including loading and splitting audio files.

    Attributes
    ----------
    sr : int
        The sample rate to use when loading audio files.
    duration : float
        The duration (in seconds) to split the audio files into segments.

    Methods
    -------
    load_audio(audio_path):
        Loads an audio file using the specified sample rate.

    split_audio(y):
        Splits the loaded audio into segments of the specified duration.
    """

    def __init__(self, sr, duration):
        """
        Constructs all the necessary attributes for the AudioProcessor object.

        Parameters
        ----------
        sr : int
            The sample rate to use when loading audio files.
        duration : float
            The duration (in seconds) to split the audio files into segments.
        """
        self.sr = sr
        self.duration = duration

    def load_audio(self, audio_path):
        """
        Loads an audio file using the specified sample rate.

        Parameters
        ----------
        audio_path : str
            The path to the audio file to be loaded.

        Returns
        -------
        tuple
            A tuple (y, sr) where y is the audio time series and sr is the sampling rate.
        """
        return librosa.load(audio_path, sr=self.sr)

    def split_audio(self, y):
        """
        Splits the loaded audio into segments of the specified duration.

        Parameters
        ----------
        y : np.ndarray
            The audio time series data.

        Returns
        -------
        list
            A list of audio segments where each segment is a numpy array.
        """
        n_samples = int(self.sr * self.duration)
        return [y[i:i + n_samples] for i in range(0, len(y), n_samples)]

단위 테스트 파일

import unittest
from unittest.mock import patch

import numpy as np

from preprocessing.audio_processor import AudioProcessor


class TestAudioProcessor(unittest.TestCase):

    def setUp(self):
        self.sr = 22050
        self.duration = 2.0
        self.processor = AudioProcessor(self.sr, self.duration)

        t = np.linspace(0, self.duration, int(self.sr * self.duration), endpoint=False)
        self.test_tone = 0.5 * np.sin(2 * np.pi * 440 * t)

    @patch('librosa.load')
    def test_load_audio(self, mock_load):
        mock_load.return_value = (self.test_tone, self.sr)

        y, sr = self.processor.load_audio("dummy_path.wav")

        mock_load.assert_called_once_with('dummy_path.wav', sr=self.sr)
        self.assertEqual(sr, self.sr)
        np.testing.assert_array_equal(y, self.test_tone)

    def test_split_audio(self):
        y = self.test_tone
        segments = self.processor.split_audio(y)
        expected_segments = int(np.ceil(len(y) / (self.sr * self.duration)))

        self.assertEqual(len(segments), expected_segments)

        for segment in segments[:-1]:
            self.assertEqual(len(segment), int(self.sr * self.duration))

        self.assertTrue(len(segments[-1]) <= int(self.sr * self.duration))


if __name__ == '__main__':
    unittest.main()

2.만들 폴더의 경로와 폴더에서 wav파일을 가져오는 클래스

import os

class FileSystemHelper:
    """
    A helper class for filesystem operations related to WAV files.

    Methods
    -------
    make_dir(path):
        Creates a directory if it doesn't already exist.

    get_wav_files(folder_path):
        Retrieves a list of paths to WAV files within a specified folder.

    """

    @staticmethod
    def make_dir(path):
        """
        Creates a directory if it doesn't already exist.

        Parameters
        ----------
        path : str
            The path of the directory to be created.
        """
        if not os.path.exists(path):
            os.makedirs(path)

    @staticmethod
    def get_wav_files(folder_path):
        """
        Retrieves a list of paths to WAV files within a specified folder.

        Parameters
        ----------
        folder_path : str
            The path of the folder where WAV files are located.

        Returns
        -------
        list
            A list of paths to WAV files (absolute paths).
        """
        return [os.path.join(folder_path, file) for file in os.listdir(folder_path) if file.endswith('wav')]

테스트 코드

import unittest
import os
from tempfile import TemporaryDirectory
from preprocessing.file_system_helper import FileSystemHelper


class TestFileSystemHelper(unittest.TestCase):

    def test_make_dir(self):
        """Test the make_dir method."""
        with TemporaryDirectory() as tempdir:
            test_path = os.path.join(tempdir, 'test_directory')
            FileSystemHelper.make_dir(test_path)

            # Verify that the directory was created
            self.assertTrue(os.path.exists(test_path))
            self.assertTrue(os.path.isdir(test_path))

    def test_get_wav_files(self):
        """Test the get_wav_files method."""
        with TemporaryDirectory() as tempdir:
            test_folder = os.path.join(tempdir, 'test_folder')
            os.makedirs(test_folder, exist_ok=True)

            # Create test files
            test_files = ['test1.wav', 'test2.wav', 'test3.txt']
            for filename in test_files:
                open(os.path.join(test_folder, filename), 'a').close()

            # Call the method under test
            wav_files = FileSystemHelper.get_wav_files(test_folder)

            # Verify the result
            expected_files = [os.path.join(test_folder, file) for file in ['test1.wav', 'test2.wav']]
            self.assertEqual(sorted(wav_files), sorted(expected_files))


if __name__ == '__main__':
    unittest.main()

3.단일 wav 음원을 mel-spectrogram으로 바꾸는 클래스

import librosa
import matplotlib.pyplot as plt
import numpy as np
import librosa.display

from preprocessing.file_system_helper import FileSystemHelper


class MelSpectrogramGenerator:
    """
    A class to generate and save Mel spectrograms from audio data.

    Attributes
    ----------
    save_path : str
        The path where the generated spectrograms will be saved.
    sr : int
        The sample rate of the audio data.
    n_mels : int
        Number of Mel bands to generate.
    hop_length : int
        Hop length (in samples) between successive frames.

    Methods
    -------
    generate_and_save(y, file_name):
        Generates a Mel spectrogram from audio data `y` and saves it as an image.

    """

    def __init__(self, save_path, sr, n_fft, n_mels, hop_length):
        """
        Initializes the MelSpectrogramGenerator instance.

        Parameters
        ----------
        save_path : str
            The path where the generated spectrograms will be saved.
        sr : int
            The sample rate of the audio data.
        n_mels : int
            Number of Mel bands to generate.
        hop_length : int
            Hop length (in samples) between successive frames.
        """
        self.save_path = save_path
        self.sr = sr
        self.n_fft = n_fft
        self.n_mels = n_mels
        self.hop_length = hop_length
        FileSystemHelper.ensure_directory_exists(save_path)

    def generate_and_save(self, y, file_name):
        """
        Generates a Mel spectrogram from audio data `y` and saves it as an image.

        Parameters
        ----------
        y : np.ndarray
            Audio time series.
        file_name : str
            File name to save the generated spectrogram image.
        """
        # Generate Mel spectrogram and convert to log scale
        mel = librosa.feature.melspectrogram(y=y, sr=self.sr,n_fft=self.n_fft, n_mels=self.n_mels, hop_length=self.hop_length)
        log_mel = librosa.power_to_db(mel, ref=np.max)

        # Plot and save the spectrogram as an image
        plt.figure(figsize=(20, 10))
        librosa.display.specshow(log_mel, sr=self.sr, hop_length=self.hop_length, cmap='viridis', x_axis='time',
                                 y_axis='mel')
        plt.colorbar(format='%+2.0f dB')  # 색상 막대 추가
        plt.savefig(f"{self.save_path}/{file_name}", bbox_inches='tight', pad_inches=0)
        plt.close()

테스트 코드

import unittest
import numpy as np
from unittest.mock import patch, MagicMock
from preprocessing.mel_spectrogram_generator import MelSpectrogramGenerator
import tempfile
import os
import warnings



class TestMelSpectrogramGenerator(unittest.TestCase):

    def setUp(self):
        """Set up the test case."""
        self.temp_dir = tempfile.TemporaryDirectory()
        self.save_path = self.temp_dir.name  # Use temporary directory
        self.sr = 22050
        self.n_mels = 128
        self.hop_length = 512

        # Create a MagicMock for librosa.display.specshow
        self.mock_specshow = MagicMock()

        # Create a MagicMock for matplotlib.pyplot.savefig
        self.mock_savefig = MagicMock()

        self.generator = MelSpectrogramGenerator(self.save_path, self.sr, self.n_mels, self.hop_length)

    def tearDown(self):
        """Clean up the temporary directory."""
        self.temp_dir.cleanup()

    @patch('librosa.feature.melspectrogram')
    @patch('librosa.power_to_db')
    @patch('librosa.display.specshow', new_callable=MagicMock)
    @patch('matplotlib.pyplot.savefig', new_callable=MagicMock)
    def test_generate_and_save(self, mock_savefig, mock_specshow, mock_power_to_db, mock_melspectrogram):
        """Test the generate_and_save method."""
        # Mock return values
        mock_mel = np.zeros((self.n_mels, 100))  # Mock Mel spectrogram
        mock_melspectrogram.return_value = mock_mel  # Ensure consistent shapes
        mock_power_to_db.return_value = mock_mel

        # Call the method under test
        test_y = np.random.randn(1000)
        test_file_name = 'test_spec'
        self.generator.generate_and_save(test_y, test_file_name)

        # Assertions to verify the behavior
        mock_melspectrogram.assert_called_once_with(test_y, sr=self.sr, n_mels=self.n_mels, hop_length=self.hop_length)
        mock_power_to_db.assert_called_once_with(mock_mel, ref=np.max)
        mock_specshow.assert_called_once()
        mock_savefig.assert_called_once_with(f"{self.save_path}/{test_file_name}.png", bbox_inches='tight',
                                             pad_inches=0)


if __name__ == '__main__':
    unittest.main()

4.1, 2, 3을 모두 합쳐 폴더 안의 wav 파일을 mel-spectrogram으로 바꾸는 클래스

class ConvertToMelSpectrogram:
    """
    A class to convert audio files in a folder to mel spectrogram images.

    Attributes:
        folder_path (str): Path to the folder containing audio files.
        save_path (str): Path to the folder where spectrogram images will be saved.
        sr (int): Sample rate for audio processing.
        duration (float): Duration of the audio clips to be processed.
        n_mels (int): Number of mel bands to generate.
        hop_length (int): Number of samples between successive frames.
        audio_processor (AudioProcessor): Instance of AudioProcessor to handle audio loading and processing.
        mel_generator (MelSpectrogramGenerator): Instance of MelSpectrogramGenerator to generate and save mel spectrograms.
    """

    def __init__(self, folder_path, save_path, sr, n_fft, duration, n_mels, hop_length):
        """
        Initializes the ConverToMelSpectrogram class with provided parameters.

        Args:
            folder_path (str): Path to the folder containing audio files.
            save_path (str): Path to the folder where spectrogram images will be saved.
            sr (int): Sample rate for audio processing.
            duration (float): Duration of the audio clips to be processed.
            n_mels (int): Number of mel bands to generate.
            hop_length (int): Number of samples between successive frames.
        """
        self.folder_path = folder_path
        self.save_path = save_path
        self.sr = sr
        self.n_fft = n_fft
        self.duration = duration
        self.n_mels = n_mels
        self.hop_length = hop_length

        self.audio_processor = AudioProcessor(sr, duration)
        self.mel_generator = MelSpectrogramGenerator(save_path, sr, n_fft, n_mels, hop_length)

    def convert_folder_to_mel_spectrogram(self):
        """
        Converts all audio files in the specified folder to mel spectrogram images.
        The spectrogram images are saved in the specified save_path directory.
        """
        wav_files = FileSystemHelper.get_files_by_extension(self.folder_path, 'wav')
        for audio_path in wav_files:
            y, sr = self.audio_processor.load_audio(audio_path)
            audio_slices = self.audio_processor.split_audio(y)
            for i, y_slice in enumerate(audio_slices):
                file_name = f"{os.path.splitext(os.path.basename(audio_path))[0]}_part{i}.png"
                self.mel_generator.generate_and_save(y_slice, file_name)

테스트 코드

import unittest
import os
import tempfile
import numpy as np
import soundfile as sf
import librosa
from unittest.mock import patch
from preprocessing.convert_to_mel_spectrogram import ConvertToMelSpectrogram
from preprocessing.audio_processor import AudioProcessor
from preprocessing.mel_spectrogram_generator import MelSpectrogramGenerator

class TestConvertToMelSpectrogram(unittest.TestCase):

    def setUp(self):
        self.temp_dir = tempfile.mkdtemp()  # 임시 디렉토리 생성
        self.audio_files = [
            self.create_temp_audio_file('audio1.wav', duration=5, sr=22050),
            self.create_temp_audio_file('audio2.wav', duration=3, sr=22050),
            self.create_temp_audio_file('audio3.wav', duration=7, sr=22050)
        ]
        self.converter = ConvertToMelSpectrogram(
            self.temp_dir,  # 임시 디렉토리 사용
            self.temp_dir,  # 저장 디렉토리도 임시 디렉토리로 설정
            sr=22050,
            duration=5.0,
            n_mels=128,
            hop_length=512
        )

    def tearDown(self):
        # 테스트 종료 후 임시 파일 삭제
        for audio_file in self.audio_files:
            os.remove(audio_file)

    def create_temp_audio_file(self, filename, duration, sr):
        # 임시 WAV 파일 생성
        audio_path = os.path.join(self.temp_dir, filename)
        y = np.random.randn(int(duration * sr))
        sf.write(audio_path, y, sr)  # soundfile로 WAV 파일 생성
        return audio_path

    @patch('librosa.load')
    @patch.object(MelSpectrogramGenerator, 'generate_and_save')
    def test_convert_folder_to_mel_spectrogram(self, mock_generate_and_save, mock_librosa_load):
        # 모의 처리된 librosa.load 함수 설정
        mock_librosa_load.side_effect = self.mock_librosa_load

        # convert_folder_to_mel_spectrogram 메서드 호출
        self.converter.convert_folder_to_mel_spectrogram()

        # generate_and_save가 적절한 파일명으로 호출되었는지 확인
        expected_calls = [
            unittest.mock.call(np.array([0.0]), 'audio1_part0.png'),
            unittest.mock.call(np.array([0.0]), 'audio2_part0.png'),
            unittest.mock.call(np.array([0.0]), 'audio3_part0.png'),
        ]
        mock_generate_and_save.assert_has_calls(expected_calls, any_order=True)

    def mock_librosa_load(self, audio_path, sr=None):
        # 임의의 오디오 데이터 반환
        return np.array([0.0]), 22050

if __name__ == '__main__':
    unittest.main()

실제 main 파일

메인 파일을 2개를 만들어서 관리하자.
preprocessing_main과 model_main을 만들어서 실제로 실행할 것 이다.

import sys
import os

sys.path.append(os.path.join(os.path.dirname(__file__), 'src'))

from preprocessing.convert_to_mel_spectrogram import ConvertToMelSpectrogram

def main():
    folder_path = '/Users/seong-gyeongjun/Downloads/vocal artist/stone/wav'
    save_path = '/Users/seong-gyeongjun/Downloads/vocal artist/stone/mel'

    sr = 22050
    n_fft = 2048
    duration = 3.0
    n_mels = 128
    hop_length = n_fft // 4

    converter = ConvertToMelSpectrogram(folder_path, save_path, sr, n_fft, duration, n_mels, hop_length)

    converter.convert_folder_to_mel_spectrogram()

if __name__ == '__main__':
    main()

결과물

이런 이미지 파일이 save_folder에 수백장 저장된 걸 볼 수 있다.