Audio Datasets

Author: Moto Hira

torchaudio provides easy access to common, publicly accessible datasets. Please refer to the official documentation for the list of available datasets.

import torch
import torchaudio

print(torch.__version__)
print(torchaudio.__version__)

2.1.1
2.1.2

import os

import IPython

import matplotlib.pyplot as plt


_SAMPLE_DIR = "_assets"
YESNO_DATASET_PATH = os.path.join(_SAMPLE_DIR, "yes_no")
os.makedirs(YESNO_DATASET_PATH, exist_ok=True)


def plot_specgram(waveform, sample_rate, title="Spectrogram"):
    waveform = waveform.numpy()

    figure, ax = plt.subplots()
    ax.specgram(waveform[0], Fs=sample_rate)
    figure.suptitle(title)
    figure.tight_layout()

Here, we show how to use the torchaudio.datasets.YESNO dataset.

dataset = torchaudio.datasets.YESNO(YESNO_DATASET_PATH, download=True)

  0%|          | 0.00/4.49M [00:00<?, ?B/s]
  1%|          | 40.0k/4.49M [00:00<00:20, 227kB/s]
  5%|4         | 208k/4.49M [00:00<00:06, 646kB/s]
 19%|#8        | 864k/4.49M [00:00<00:01, 2.02MB/s]
 60%|######    | 2.70M/4.49M [00:00<00:00, 5.40MB/s]
100%|##########| 4.49M/4.49M [00:00<00:00, 5.67MB/s]

i = 1
waveform, sample_rate, label = dataset[i]
plot_specgram(waveform, sample_rate, title=f"Sample {i}: {label}")
IPython.display.Audio(waveform, rate=sample_rate)

i = 3
waveform, sample_rate, label = dataset[i]
plot_specgram(waveform, sample_rate, title=f"Sample {i}: {label}")
IPython.display.Audio(waveform, rate=sample_rate)

i = 5
waveform, sample_rate, label = dataset[i]
plot_specgram(waveform, sample_rate, title=f"Sample {i}: {label}")
IPython.display.Audio(waveform, rate=sample_rate)