audio-classification

Dataset setup

Use the helper script in scripts/download_data.py to fetch the Audio Scene Classification dataset and unpack it into data/raw/audiosceneclassification2025.

python scripts/download_data.py

The script is idempotent—it will reuse the existing archive or extracted folder unless you pass --force-download and/or --force-extract. The downloaded archive is saved under data/raw/audiosceneclassification2025.zip.

Listen to a sample clip

Preview a random clip (or specify one) along with its label:

python scripts/listen_random_clip.py
# or specify a particular clip
python scripts/listen_random_clip.py --clip a001_50_60.wav

The script looks inside data/raw/audiosceneclassification2025/audio and uses the metadata in meta.txt to show the label/clip ID.

Training a model

we cannot feed an audio file .wav directly to a model. We need to convert it into a tensor first (waveform).

waveform, sr = torchaudio.load("audio.wav")
waveform.shape  # torch.Size([num_channels, num_samples]) the number of channels and number of samples on each channel

print(waveform[:, :10])   # first 10 samples of each channel
tensor([[ 0.12, 0.50, 0.80, 0.20, -0.10, -0.50, -0.80, -0.30, 0.10, 0.40]])

For consistency, we will convert to mono channel (single channel) for any audio channels >1 Then we will pad or crop the waveform to a fixed length (e.g., 10 seconds) to ensure consistent input size for the model. (Maybe not needed if all audio files are already of the same time length. But for unexpected cases or for audio augmentation later, it's good to have this step.)