音频数据集

作者: Moto Hira

torchaudio 提供了便捷的方式来访问常见且公开可用的数据集。请参阅官方文档以获取可用数据集的列表。

import torch
import torchaudio

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

2.6.0.dev20241104
2.5.0.dev20241105

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()

此处，我们展示如何使用 torchaudio.datasets.YESNO 数据集。

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

  0%|          | 0.00/4.49M [00:00<?, ?B/s]
  3%|2         | 128k/4.49M [00:00<00:08, 514kB/s]
 11%|#1        | 512k/4.49M [00:00<00:03, 1.39MB/s]
 39%|###9      | 1.75M/4.49M [00:00<00:00, 4.44MB/s]
 81%|########  | 3.62M/4.49M [00:00<00:00, 8.03MB/s]
100%|##########| 4.49M/4.49M [00:00<00:00, 6.89MB/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)