注意
点击 这里 下载完整示例代码
音频重采样¶
在这里，我们将逐步介绍如何使用 torchaudio 对音频波形进行重采样。
# When running this tutorial in Google Colab, install the required packages
# with the following.
# !pip install torchaudio librosa
import torch
import torchaudio
import torchaudio.functional as F
import torchaudio.transforms as T
print(torch.__version__)
print(torchaudio.__version__)
Out:
1.11.0+cpu
0.11.0+cpu
准备数据和实用函数（跳过此部分）¶
# @title Prepare data and utility functions. {display-mode: "form"}
# @markdown
# @markdown You do not need to look into this cell.
# @markdown Just execute once and you are good to go.
# -------------------------------------------------------------------------------
# Preparation of data and helper functions.
# -------------------------------------------------------------------------------
import math
import time
import librosa
import matplotlib.pyplot as plt
import pandas as pd
from IPython.display import Audio, display
DEFAULT_OFFSET = 201
SWEEP_MAX_SAMPLE_RATE = 48000
DEFAULT_LOWPASS_FILTER_WIDTH = 6
DEFAULT_ROLLOFF = 0.99
DEFAULT_RESAMPLING_METHOD = "sinc_interpolation"
def _get_log_freq(sample_rate, max_sweep_rate, offset):
"""Get freqs evenly spaced out in log-scale, between [0, max_sweep_rate // 2]
offset is used to avoid negative infinity `log(offset + x)`.
"""
start, stop = math.log(offset), math.log(offset + max_sweep_rate // 2)
return torch.exp(torch.linspace(start, stop, sample_rate, dtype=torch.double)) - offset
def _get_inverse_log_freq(freq, sample_rate, offset):
"""Find the time where the given frequency is given by _get_log_freq"""
half = sample_rate // 2
return sample_rate * (math.log(1 + freq / offset) / math.log(1 + half / offset))
def _get_freq_ticks(sample_rate, offset, f_max):
# Given the original sample rate used for generating the sweep,
# find the x-axis value where the log-scale major frequency values fall in
time, freq = [], []
for exp in range(2, 5):
for v in range(1, 10):
f = v * 10 ** exp
if f < sample_rate // 2:
t = _get_inverse_log_freq(f, sample_rate, offset) / sample_rate
time.append(t)
freq.append(f)
t_max = _get_inverse_log_freq(f_max, sample_rate, offset) / sample_rate
time.append(t_max)
freq.append(f_max)
return time, freq
def get_sine_sweep(sample_rate, offset=DEFAULT_OFFSET):
max_sweep_rate = sample_rate
freq = _get_log_freq(sample_rate, max_sweep_rate, offset)
delta = 2 * math.pi * freq / sample_rate
cummulative = torch.cumsum(delta, dim=0)
signal = torch.sin(cummulative).unsqueeze(dim=0)
return signal
def plot_sweep(
waveform,
sample_rate,
title,
max_sweep_rate=SWEEP_MAX_SAMPLE_RATE,
offset=DEFAULT_OFFSET,
):
x_ticks = [100, 500, 1000, 5000, 10000, 20000, max_sweep_rate // 2]
y_ticks = [1000, 5000, 10000, 20000, sample_rate // 2]
time, freq = _get_freq_ticks(max_sweep_rate, offset, sample_rate // 2)
freq_x = [f if f in x_ticks and f <= max_sweep_rate // 2 else None for f in freq]
freq_y = [f for f in freq if f >= 1000 and f in y_ticks and f <= sample_rate // 2]
figure, axis = plt.subplots(1, 1)
axis.specgram(waveform[0].numpy(), Fs=sample_rate)
plt.xticks(time, freq_x)
plt.yticks(freq_y, freq_y)
axis.set_xlabel("Original Signal Frequency (Hz, log scale)")
axis.set_ylabel("Waveform Frequency (Hz)")
axis.xaxis.grid(True, alpha=0.67)
axis.yaxis.grid(True, alpha=0.67)
figure.suptitle(f"{title} (sample rate: {sample_rate} Hz)")
plt.show(block=True)
def play_audio(waveform, sample_rate):
waveform = waveform.numpy()
num_channels, num_frames = waveform.shape
if num_channels == 1:
display(Audio(waveform[0], rate=sample_rate))
elif num_channels == 2:
display(Audio((waveform[0], waveform[1]), rate=sample_rate))
else:
raise ValueError("Waveform with more than 2 channels are not supported.")
def plot_specgram(waveform, sample_rate, title="Spectrogram", xlim=None):
waveform = waveform.numpy()
num_channels, num_frames = waveform.shape
figure, axes = plt.subplots(num_channels, 1)
if num_channels == 1:
axes = [axes]
for c in range(num_channels):
axes[c].specgram(waveform[c], Fs=sample_rate)
if num_channels > 1:
axes[c].set_ylabel(f"Channel {c+1}")
if xlim:
axes[c].set_xlim(xlim)
figure.suptitle(title)
plt.show(block=False)
def benchmark_resample(
method,
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=DEFAULT_LOWPASS_FILTER_WIDTH,
rolloff=DEFAULT_ROLLOFF,
resampling_method=DEFAULT_RESAMPLING_METHOD,
beta=None,
librosa_type=None,
iters=5,
):
if method == "functional":
begin = time.time()
for _ in range(iters):
F.resample(
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=lowpass_filter_width,
rolloff=rolloff,
resampling_method=resampling_method,
)
elapsed = time.time() - begin
return elapsed / iters
elif method == "transforms":
resampler = T.Resample(
sample_rate,
resample_rate,
lowpass_filter_width=lowpass_filter_width,
rolloff=rolloff,
resampling_method=resampling_method,
dtype=waveform.dtype,
)
begin = time.time()
for _ in range(iters):
resampler(waveform)
elapsed = time.time() - begin
return elapsed / iters
elif method == "librosa":
waveform_np = waveform.squeeze().numpy()
begin = time.time()
for _ in range(iters):
librosa.resample(waveform_np, orig_sr=sample_rate, target_sr=resample_rate, res_type=librosa_type)
elapsed = time.time() - begin
return elapsed / iters
重采样概述¶
要将音频波形从一个频率重新采样到另一个频率，可以使用
torchaudio.transforms.Resample() 或
torchaudio.functional.resample()。
transforms.Resample 会预先计算并缓存用于重采样的内核，
而 functional.resample 则是在需要时动态计算，因此使用
torchaudio.transforms.Resample 在使用相同参数对多个音频波形进行重采样时会提高速度（参见基准测试部分）。
两种重采样方法都使用带限sinc插值来计算任意时间步长的信号值。实现涉及卷积，因此我们可以利用GPU / 多线程来提高性能。在多个子进程中使用重采样时，例如使用多个工作进程进行数据加载，您的应用程序可能会创建比系统能够有效处理更多的线程。在这种情况下，设置torch.set_num_threads(1)可能会有所帮助。
由于有限数量的样本只能表示有限数量的频率，重采样并不能产生完美的结果，可以通过多种参数来控制其质量和计算速度。我们通过重采样一个对数正弦扫频来演示这些特性，这是一种随时间呈指数增长频率的正弦波。
下图中的频谱图显示了信号的频率表示， 其中 x 轴对应原始波形的频率（以对数刻度表示）， y 轴对应所绘制波形的频率， 颜色强度表示振幅。
sample_rate = 48000
resample_rate = 32000
waveform = get_sine_sweep(sample_rate)
plot_sweep(waveform, sample_rate, title="Original Waveform")
play_audio(waveform, sample_rate)
resampler = T.Resample(sample_rate, resample_rate, dtype=waveform.dtype)
resampled_waveform = resampler(waveform)
plot_sweep(resampled_waveform, resample_rate, title="Resampled Waveform")
play_audio(waveform, sample_rate)
Out:
<IPython.lib.display.Audio object>
<IPython.lib.display.Audio object>
通过参数控制重采样质量¶
低通滤波器宽度¶
因为用于插值的滤波器是无限延伸的，所以使用
lowpass_filter_width 参数来控制用于窗口插值的滤波器宽度。它也被称为零交叉次数，因为插值在每个时间单位都会穿过零点。使用更大的 lowpass_filter_width 可以提供更尖锐、更精确的滤波器，但计算成本更高。
sample_rate = 48000
resample_rate = 32000
resampled_waveform = F.resample(waveform, sample_rate, resample_rate, lowpass_filter_width=6)
plot_sweep(resampled_waveform, resample_rate, title="lowpass_filter_width=6")
resampled_waveform = F.resample(waveform, sample_rate, resample_rate, lowpass_filter_width=128)
plot_sweep(resampled_waveform, resample_rate, title="lowpass_filter_width=128")
回滚¶
The rolloff 参数表示奈奎斯特频率的一个分数，奈奎斯特频率是给定有限采样率所能表示的最大频率。rolloff 决定了低通滤波器的截止频率，并控制混叠的程度，混叠发生在高于奈奎斯特频率的频率被映射到较低频率时。因此，较低的滚降将减少混叠量，但也同时会降低部分较高频率。
sample_rate = 48000
resample_rate = 32000
resampled_waveform = F.resample(waveform, sample_rate, resample_rate, rolloff=0.99)
plot_sweep(resampled_waveform, resample_rate, title="rolloff=0.99")
resampled_waveform = F.resample(waveform, sample_rate, resample_rate, rolloff=0.8)
plot_sweep(resampled_waveform, resample_rate, title="rolloff=0.8")
窗口函数¶
默认情况下，torchaudio 的重采样使用 Hann 窗口滤波器，这是一个加权余弦函数。它还支持 Kaiser 窗口，这是一种接近最优的窗口函数，包含一个额外的
beta 参数，可用于设计滤波器的平滑度和脉冲宽度。这可以通过使用
resampling_method 参数进行控制。
sample_rate = 48000
resample_rate = 32000
resampled_waveform = F.resample(waveform, sample_rate, resample_rate, resampling_method="sinc_interpolation")
plot_sweep(resampled_waveform, resample_rate, title="Hann Window Default")
resampled_waveform = F.resample(waveform, sample_rate, resample_rate, resampling_method="kaiser_window")
plot_sweep(resampled_waveform, resample_rate, title="Kaiser Window Default")
与librosa的对比¶
torchaudio’s resample 函数可以用来生成与 librosa (resampy) 的 Kaiser 窗重采样类似的结果，但会带有一些噪声
sample_rate = 48000
resample_rate = 32000
# kaiser_best
resampled_waveform = F.resample(
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=64,
rolloff=0.9475937167399596,
resampling_method="kaiser_window",
beta=14.769656459379492,
)
plot_sweep(resampled_waveform, resample_rate, title="Kaiser Window Best (torchaudio)")
librosa_resampled_waveform = torch.from_numpy(
librosa.resample(waveform.squeeze().numpy(), orig_sr=sample_rate, target_sr=resample_rate, res_type="kaiser_best")
).unsqueeze(0)
plot_sweep(librosa_resampled_waveform, resample_rate, title="Kaiser Window Best (librosa)")
mse = torch.square(resampled_waveform - librosa_resampled_waveform).mean().item()
print("torchaudio and librosa kaiser best MSE:", mse)
# kaiser_fast
resampled_waveform = F.resample(
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=16,
rolloff=0.85,
resampling_method="kaiser_window",
beta=8.555504641634386,
)
plot_specgram(resampled_waveform, resample_rate, title="Kaiser Window Fast (torchaudio)")
librosa_resampled_waveform = torch.from_numpy(
librosa.resample(waveform.squeeze().numpy(), orig_sr=sample_rate, target_sr=resample_rate, res_type="kaiser_fast")
).unsqueeze(0)
plot_sweep(librosa_resampled_waveform, resample_rate, title="Kaiser Window Fast (librosa)")
mse = torch.square(resampled_waveform - librosa_resampled_waveform).mean().item()
print("torchaudio and librosa kaiser fast MSE:", mse)
Out:
torchaudio and librosa kaiser best MSE: 2.0806901153659873e-06
torchaudio and librosa kaiser fast MSE: 2.5200744248601027e-05
性能基准测试¶
以下是两个采样率对之间下采样和上采样波形的基准测试。我们展示了lowpass_filter_wdith、窗口类型和采样率可能带来的性能影响。此外，我们还提供了与librosa的kaiser_best和kaiser_fast的对比，使用它们在torchaudio中的相应参数。
进一步阐述结果：
更大的
lowpass_filter_width会导致更大的重采样内核， 因此会增加内核计算和卷积的计算时间使用
kaiser_window会导致比默认值sinc_interpolation更长的计算时间，因为计算中间窗口值更为复杂——采样率与重采样率之间较大的最大公约数（GCD）将导致简化，从而允许使用更小的核并加快核计算速度。
configs = {
"downsample (48 -> 44.1 kHz)": [48000, 44100],
"downsample (16 -> 8 kHz)": [16000, 8000],
"upsample (44.1 -> 48 kHz)": [44100, 48000],
"upsample (8 -> 16 kHz)": [8000, 16000],
}
for label in configs:
times, rows = [], []
sample_rate = configs[label][0]
resample_rate = configs[label][1]
waveform = get_sine_sweep(sample_rate)
# sinc 64 zero-crossings
f_time = benchmark_resample("functional", waveform, sample_rate, resample_rate, lowpass_filter_width=64)
t_time = benchmark_resample("transforms", waveform, sample_rate, resample_rate, lowpass_filter_width=64)
times.append([None, 1000 * f_time, 1000 * t_time])
rows.append("sinc (width 64)")
# sinc 6 zero-crossings
f_time = benchmark_resample("functional", waveform, sample_rate, resample_rate, lowpass_filter_width=16)
t_time = benchmark_resample("transforms", waveform, sample_rate, resample_rate, lowpass_filter_width=16)
times.append([None, 1000 * f_time, 1000 * t_time])
rows.append("sinc (width 16)")
# kaiser best
lib_time = benchmark_resample("librosa", waveform, sample_rate, resample_rate, librosa_type="kaiser_best")
f_time = benchmark_resample(
"functional",
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=64,
rolloff=0.9475937167399596,
resampling_method="kaiser_window",
beta=14.769656459379492,
)
t_time = benchmark_resample(
"transforms",
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=64,
rolloff=0.9475937167399596,
resampling_method="kaiser_window",
beta=14.769656459379492,
)
times.append([1000 * lib_time, 1000 * f_time, 1000 * t_time])
rows.append("kaiser_best")
# kaiser fast
lib_time = benchmark_resample("librosa", waveform, sample_rate, resample_rate, librosa_type="kaiser_fast")
f_time = benchmark_resample(
"functional",
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=16,
rolloff=0.85,
resampling_method="kaiser_window",
beta=8.555504641634386,
)
t_time = benchmark_resample(
"transforms",
waveform,
sample_rate,
resample_rate,
lowpass_filter_width=16,
rolloff=0.85,
resampling_method="kaiser_window",
beta=8.555504641634386,
)
times.append([1000 * lib_time, 1000 * f_time, 1000 * t_time])
rows.append("kaiser_fast")
df = pd.DataFrame(times, columns=["librosa", "functional", "transforms"], index=rows)
df.columns = pd.MultiIndex.from_product([[f"{label} time (ms)"], df.columns])
display(df.round(2))
Out:
downsample (48 -> 44.1 kHz) time (ms) ...
librosa ... transforms
sinc (width 64) NaN ... 0.39
sinc (width 16) NaN ... 0.32
kaiser_best 36.50 ... 0.39
kaiser_fast 9.85 ... 0.36
[4 rows x 3 columns]
downsample (16 -> 8 kHz) time (ms) ...
librosa ... transforms
sinc (width 64) NaN ... 0.78
sinc (width 16) NaN ... 0.33
kaiser_best 12.68 ... 0.78
kaiser_fast 4.53 ... 0.34
[4 rows x 3 columns]
upsample (44.1 -> 48 kHz) time (ms) ...
librosa ... transforms
sinc (width 64) NaN ... 0.41
sinc (width 16) NaN ... 0.37
kaiser_best 33.73 ... 0.43
kaiser_fast 10.05 ... 0.42
[4 rows x 3 columns]
upsample (8 -> 16 kHz) time (ms) ...
librosa ... transforms
sinc (width 64) NaN ... 0.36
sinc (width 16) NaN ... 0.21
kaiser_best 22.68 ... 0.41
kaiser_fast 5.27 ... 0.18
[4 rows x 3 columns]
脚本的总运行时间： ( 0 分钟 4.470 秒)