Pytorch音频处理
原文:https://pytorch.org/tutorials/beginner/audio_preprocessing_tutorial.html
Pytorch Audio Processing使用torchaudio这个库。
import matplotlib.pyplot as pltimport torchaudio
打开一个音频文件
# 音频文件的双声道很接近,所以图上波形边缘有不太明显的两个颜色。filename = 'data/diarizationExample_sr16k_ac2.wav'waveform, sample_rate = torchaudio.load(filename)print("Shape of waveform: {}".format(waveform.size()))print("Sample rate of waveform: {}".format(sample_rate))plt.figure()plt.plot(waveform.t().numpy())plt.show()Shape of waveform: torch.Size([2, 672096])Sample rate of waveform: 16000
转换(Transformations)
torchaudio支持的转换列表还在增长中,这里查看。
- 重采样(Resample): Resample waveform to a different sample rate.
- 频谱图(Spectrogram): Create a spectrogram from a waveform.
- 梅尔刻度(MelScale): This turns a normal STFT into a Mel-frequency STFT, using a conversion matrix.
- 振幅转分贝(AmplitudeToDB): This turns a spectrogram from the power/amplitude scale to the decibel scale.
- 梅尔频率倒谱系数(MFCC): Create the Mel-frequency cepstrum coefficients from a waveform.
- 梅尔频谱图(MelSpectrogram): Create MEL Spectrograms from a waveform using the STFT function in PyTorch.
- μ-law编码(MuLawEncoding): Encode waveform based on mu-law companding. 原理与为何增加SNR参考这篇:https://www.mahong.me/archives/13
- μ-law解码(MuLawDecoding): Decode mu-law encoded waveform.
首先,在对数刻度上查看频谱图的对数。
specgram = torchaudio.transforms.Spectrogram()(waveform)# 双声道的音频print("Shape of spectrogram: {}".format(specgram.size()))plt.figure()plt.imshow(specgram.log2()[0,:,:].numpy())plt.show()Shape of spectrogram: torch.Size([2, 201, 3361])
或者,可以以对数刻度查看梅尔频谱图。
specgram = torchaudio.transforms.MelSpectrogram()(waveform)print("Shape of spectrogram: {}".format(specgram.size()))plt.figure()# MelSpectrogram的接口返回的跟其它几个不一样,还得detach()生成一个不需要求导的张量。# tensor.detach() creates a tensor that shares storage with tensor that does not require grad.# Ref: https://discuss.pytorch.org/t/clone-and-detach-in-v0-4-0/16861/2plt.imshow(specgram.detach()[0, :, :].numpy())plt.show()Shape of spectrogram: torch.Size([2, 128, 3361])
可以对音频进行频率重采样,一次操作一个声道。
new_sample_rate = sample_rate/10# 这里对第一个声道进行重采样channel = 0# view(1, -1)是将waveform[channel,:]这个一维数组重新组装为二维数组。transformed = torchaudio.transforms.Resample(sample_rate, new_sample_rate)(waveform[channel,:].view(1, -1))print("Shape of transformed waveform: {}".format(transformed.size()))plt.figure()plt.plot(transformed[0, :].numpy())plt.show()Shape of transformed waveform: torch.Size([1, 67210])
μ-law编码。μ-law编码要求信号的数值在-1与1之间,因为上面生成的waveform张量是常规的Pytorch张量,因此这里可以对其进行直接操作。
# 检查信号数值是否在区间[-1,1]中print("Min: {}Max: {}Mean: {}".format(waveform.min(), waveform.max(), waveform.mean()))Min: -0.5394287109375Max: 0.664764404296875Mean: 0.04157630726695061
如上所示,信号已经在区间[-1, 1],因此不需要再正则化。
def normalize(tensor): # 减去平均值,并缩放到区间[-1,1] tensor_minusmean = tensor - tensor.mean() return tensor_minusmean/tensor_minusmean.abs().max()# 归一化为[-1,1]# waveform = normalize(waveform)
对waveform进行μ-law编码。 一般来说语音信号是符合拉普拉斯分布的,当我们使用线性量化的时候则会造成一些不必要的量化等级的浪费。 因此,可以将信号先进行放大,使其的pdf(概率密度函数)分布发生改变,变得更加的均匀,然后再进行量化反转,从而得到最终的信号。(参考https://www.mahong.me/archives/13)
transformed = torchaudio.transforms.MuLawEncoding()(waveform)print("Shape of transformed waveform: {}".format(transformed.size()))plt.figure()plt.plot(transformed[0, :].numpy())plt.show()Shape of transformed waveform: torch.Size([2, 672096])
μ-law解码。
reconstructed = torchaudio.transforms.MuLawDecoding()(transformed)print("Shape of recovered waveform: {}".format(reconstructed.size()))plt.figure()plt.plot(reconstructed[0, :].numpy())plt.show()Shape of recovered waveform: torch.Size([2, 672096])
最后,对比原始waveform与经过μ-law编解码重建后的waveform。
# 计算中位数相对差err = ((waveform-reconstructed).abs()/waveform.abs()).median()print("Median relative difference between original and Mulaw rconstructed signals :{:.2%}".format(err))Median relative difference between original and Mulaw rconstructed signals :1.22%
从Kaldi迁移至torchaudio
Kaldi是活跃的语音识别工作集,torchaudio通过torchaudio.kaldi_io提供与其的兼容性。 实际上它可以通过以下方式从Kaldi scp、ark文件或流中读取内容:
- read_vec_int_ark
- read_vec_flt_scp
- read_vec_flt_arkfile/stream
- read_mat_scp
- read_mat_ark
torchaudio为频谱图(spectrogram)与过滤器组(fbank)提供与Kaldi兼容的转换(transform),参见此处获取更多信息。
n_fft = 400.0frame_length = n_fft / sample_rate * 1000.0frame_shift = frame_length / 2.0params = { "channel": 0, "dither": 0.0, "window_type": "hanning", "frame_length": frame_length, "frame_shift": frame_shift, "remove_dc_offset": False, "round_to_power_of_two": False, "sample_frequency": sample_rate,}specgram = torchaudio.compliance.kaldi.spectrogram(waveform, **params)print("Shape of spectrogram: {}".format(specgram.size()))plt.figure()plt.imshow(specgram.t().numpy())plt.show()Shape of spectrogram: torch.Size([3359, 201])
torchaudio还支持根据波形(waveform)计算滤波器组(filterbank)的功能,以匹配Kaldi的实现。
fbank = torchaudio.compliance.kaldi.fbank(waveform, **params)print("Shape of fbank: {}".format(fbank.size()))plt.figure()plt.imshow(fbank.t().numpy())plt.show()Shape of fbank: torch.Size([3359, 23])
小结
本文以原始音频信号(raw audio signal)或波形(waveform)为例,来说明如何使用torchaudio打开音频文件,以及如何预处理和转换(transform)此类波形。鉴于torchaudio是基于Pytorch构建的,这些技术可在利用GPU的同时作为更高级音频应用(例如语音识别)的构建块。