The document outlines the process of creating a custom audio dataset using PyTorch and Torchaudio, with the UrbanSound8K dataset as a practical example. It covers library imports, dataset class creation, audio transformations, and the extraction of mel spectrograms for audio classification tasks. The key takeaways emphasize the importance of audio processing techniques in machine learning applications.

1. Creating a Custom Audio Dataset with PyTorch
Extracting Mel Spectrograms with Pytorch and
Torchaudio: UrbanSound8K Dataset
[By KAHSAY]
August 22, 2024
1

2. Introduction
• Objective:
• Learn how to create a custom audio dataset using PyTorch
and torchaudio.
• Use the UrbanSound8K dataset as a practical example.
• Focus Areas:
• Basic I/O operations with torchaudio.
• Handling audio files and annotations.
2

3. Libraries and Imports
• Import Libraries:
• os for handling file paths.
• torch.utils.data.Dataset for creating custom
datasets.
• pandas for reading CSV files.
• torchaudio for loading and processing audio files.
1 import os
2 import torch
3 from torch.utils.data import Dataset
4 import pandas as pd
5 import torchaudio
3

4. UrbanSoundDataset Class Overview
• Purpose:
• Create a custom dataset class that handles the
UrbanSound8K dataset.
• Inherit from PyTorch’s Dataset class.
1 class UrbanSoundDataset(Dataset):
2 def __init__(self, annotations_file, audio_dir,
transformation,
3 target_sample_rate):
4 self.annotations = pd.read_csv(annotations_file)
5 self.audio_dir = audio_dir
6 self.transformation = transformation
7 self.target_sample_rate = target_sample_rate
8
9 def __len__(self):
10 return len(self.annotations)
11
12 def __getitem__(self, index):
13 audio_sample_path = self._get_audio_sample_path(
4

5. Methods
• _resample_if_necessary():
• Resamples the audio signal to the target sample rate.
• _mix_down_if_necessary():
• Mixes down stereo audio to mono if necessary.
1 def _resample_if_necessary(self, signal, sr):
2 if sr != self.target_sample_rate:
3 resampler = torchaudio.transforms.Resample(
sr, self.target_sample_rate)
4 signal = resampler(signal)
5 return signal
6
7 def _mix_down_if_necessary(self, signal):
8 if signal.shape[0] > 1:
9 signal = torch.mean(signal, dim=0, keepdim=
True)
10 return signal
5

6. Main Script Execution
• Purpose:
• Set up file paths and instantiate the
UrbanSoundDataset class.
• Print the number of samples and access the first audio
sample.
1 if __name__ == "__main__":
2 ANNOTATIONS_FILE = "/path/to/UrbanSound8K.csv"
3 AUDIO_DIR = "/path/to/audio"
4 SAMPLE_RATE = 16000
5
6 mel_spectrogram = torchaudio.transforms.
MelSpectrogram(
7 sample_rate=SAMPLE_RATE,
8 n_fft=1024,
9 hop_length=512,
10 n_mels=64
11 )
12
6

7. Extracting Mel Spectrograms
• Mel Spectrogram: - A representation of the audio signal in
the Mel frequency scale. - Useful for capturing the
characteristics of audio signals for classification tasks.
• Using torchaudio: - Create a Mel Spectrogram using
torchaudio.transforms.MelSpectrogram().
7

8. Resampling Audio
• Resampling: - Adjusting the sample rate of audio signals
to a target sample rate. - Important for ensuring
consistent processing across different audio files.
• Using torchaudio: - Use
torchaudio.transforms.Resample() to resample
audio signals.
8

9. Common torchaudio Transforms
• MelSpectrogram: - Converts audio waveforms to Mel
spectrograms.
• Spectrogram: - Converts audio waveforms to regular
spectrograms.
• AmplitudeToDB: - Converts amplitude to decibels for
better visualization.
• TimeStretch: - Stretches the audio signal in time without
altering pitch.
• PitchShift: - Shifts the pitch of the audio signal by a
specified number of semitones.
• Apply Usage:
• Combine transforms using torch.nn.Sequential for a
pipeline.
9

10. Dataset Loading Workflow
• Workflow Steps:
1. Read Annotations:
• Load metadata from the CSV file using pandas.
2. Load Audio:
• Use torchaudio.load() to read audio files.
3. Return Data:
• Output the audio signal and its label.
10

11. UrbanSound8K Dataset
• Description:
• A dataset with 8,732 labeled sound excerpts from 10
classes.
• Useful for training audio classification models.
11

12. Conclusion
• Key Takeaways:
• Understanding the process of creating a custom audio
dataset.
• Importance of audio transformations for machine learning
tasks.
• Practical implementation using torchaudio.
12