音频特征增强

作者：Moto Hira

import torch
import torchaudio
import torchaudio.transforms as T
import numpy as np

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

2.10.0.dev20251013+cu126
2.8.0a0+1d65bbe

准备

import matplotlib.pyplot as plt
from IPython.display import Audio
from torchaudio.utils import _download_asset
import torchaudio

在本教程中，我们将使用来自 VOiCES 数据集 的语音数据，该数据集在 Creative Commos BY 4.0 许可下可用。

SAMPLE_WAV_SPEECH_PATH = _download_asset("tutorial-assets/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav")


def _get_sample(path):
    return torchaudio.load(path)


def get_speech_sample():
    return _get_sample(SAMPLE_WAV_SPEECH_PATH)


def get_spectrogram(
    n_fft=400,
    win_len=None,
    hop_len=None,
    power=2.0,
):
    waveform, _ = get_speech_sample()
    spectrogram = T.Spectrogram(
        n_fft=n_fft,
        win_length=win_len,
        hop_length=hop_len,
        center=True,
        pad_mode="reflect",
        power=power,
    )
    return spectrogram(waveform)

SpecAugment

SpecAugment 是一种流行的频谱图增强技术。

torchaudio 实现 了 torchaudio.transforms.TimeStretch(), torchaudio.transforms.TimeMasking() 和 torchaudio.transforms.FrequencyMasking()。

时间拉伸

spec = get_spectrogram(power=None)
stretch = T.TimeStretch()

spec_12 = stretch(spec, overriding_rate=1.2)
spec_09 = stretch(spec, overriding_rate=0.9)

可视化

def power_to_db(S):
    S = np.asarray(S)
    return 10.0 * np.log10(np.maximum(1e-10, S))


def plot():
    def plot_spec(ax, spec, title):
        ax.set_title(title)
        ax.imshow(power_to_db(spec**2), origin="lower", aspect="auto")

    fig, axes = plt.subplots(3, 1, sharex=True, sharey=True)
    plot_spec(axes[0], torch.abs(spec_12[0]), title="Stretched x1.2")
    plot_spec(axes[1], torch.abs(spec[0]), title="Original")
    plot_spec(axes[2], torch.abs(spec_09[0]), title="Stretched x0.9")
    fig.tight_layout()


plot()

音频样本

def preview(spec, rate=16000):
    ispec = T.InverseSpectrogram()
    waveform = ispec(spec)

    return Audio(waveform[0].numpy().T, rate=rate)


preview(spec)

preview(spec_12)

preview(spec_09)

时间和频率掩码

torch.random.manual_seed(4)

time_masking = T.TimeMasking(time_mask_param=80)
freq_masking = T.FrequencyMasking(freq_mask_param=80)

spec = get_spectrogram()
time_masked = time_masking(spec)
freq_masked = freq_masking(spec)

def plot():
    def plot_spec(ax, spec, title):
        ax.set_title(title)
        ax.imshow(power_to_db(spec), origin="lower", aspect="auto")

    fig, axes = plt.subplots(3, 1, sharex=True, sharey=True)
    plot_spec(axes[0], spec[0], title="Original")
    plot_spec(axes[1], time_masked[0], title="Masked along time axis")
    plot_spec(axes[2], freq_masked[0], title="Masked along frequency axis")
    fig.tight_layout()


plot()