AudioEffector 使用方法

本教程展示了如何使用 torchaudio.io.AudioEffector 对波形张量应用各种效果和编解码器。

本教程需要 FFmpeg 库。有关详细信息，请参阅 FFmpeg 依赖。

概述

AudioEffector 结合了由 StreamWriter 和 StreamReader 提供的内存编码、解码和过滤功能。

下图展示了该过程。

import torch
import torchaudio

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

2.6.0
2.6.0

from torchaudio.io import AudioEffector, CodecConfig

import matplotlib.pyplot as plt
from IPython.display import Audio

for k, v in torchaudio.utils.ffmpeg_utils.get_versions().items():
    print(k, v)

libavcodec (60, 3, 100)
libavdevice (60, 1, 100)
libavfilter (9, 3, 100)
libavformat (60, 3, 100)
libavutil (58, 2, 100)

用法

要使用 AudioEffector，请使用 effect 和 format 实例化它，然后将波形传递给 apply() 或 stream() 方法。

effector = AudioEffector(effect=..., format=...,)

# Apply at once
applied = effector.apply(waveform, sample_rate)

apply 方法将效果和编解码器一次性应用于整个波形。因此，如果输入波形较长，且内存消耗是一个问题，可以使用 stream 方法逐块处理。

# Apply chunk by chunk
for applied_chunk = effector.stream(waveform, sample_rate):
    ...

示例

src = torchaudio.utils.download_asset("tutorial-assets/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav")
waveform, sr = torchaudio.load(src, channels_first=False)

图库

def show(effect, *, stereo=False):
    wf = torch.cat([waveform] * 2, dim=1) if stereo else waveform
    figsize = (6.4, 2.1 if stereo else 1.2)

    effector = AudioEffector(effect=effect, pad_end=False)
    result = effector.apply(wf, int(sr))

    num_channels = result.size(1)
    f, ax = plt.subplots(num_channels, 1, squeeze=False, figsize=figsize, sharex=True)
    for i in range(num_channels):
        ax[i][0].specgram(result[:, i], Fs=sr)
    f.set_tight_layout(True)

    return Audio(result.numpy().T, rate=sr)

原始代码

show(effect=None)

effector tutorial

效果

节拍

https://ffmpeg.org/ffmpeg-filters.html#atempo

show("atempo=0.7")

effector tutorial

show("atempo=1.8")

effector tutorial

高通滤波器

https://ffmpeg.org/ffmpeg-filters.html#highpass

show("highpass=frequency=1500")

effector tutorial

低通滤波器

https://ffmpeg.org/ffmpeg-filters.html#lowpass

show("lowpass=frequency=1000")

effector tutorial

全通滤波器

https://ffmpeg.org/ffmpeg-filters.html#allpass

show("allpass")

effector tutorial

带通滤波器

https://ffmpeg.org/ffmpeg-filters.html#bandpass

show("bandpass=frequency=3000")

effector tutorial

带阻滤波器

https://ffmpeg.org/ffmpeg-filters.html#bandreject

show("bandreject=frequency=3000")

effector tutorial

回声

https://ffmpeg.org/ffmpeg-filters.html#aecho

show("aecho=in_gain=0.8:out_gain=0.88:delays=6:decays=0.4")

effector tutorial

show("aecho=in_gain=0.8:out_gain=0.88:delays=60:decays=0.4")

effector tutorial

show("aecho=in_gain=0.8:out_gain=0.9:delays=1000:decays=0.3")

effector tutorial

合唱效果

https://ffmpeg.org/ffmpeg-filters.html#chorus

show("chorus=0.5:0.9:50|60|40:0.4|0.32|0.3:0.25|0.4|0.3:2|2.3|1.3")

effector tutorial

FFT 滤波器

https://ffmpeg.org/ffmpeg-filters.html#afftfilt

# fmt: off
show(
    "afftfilt="
    "real='re * (1-clip(b * (b/nb), 0, 1))':"
    "imag='im * (1-clip(b * (b/nb), 0, 1))'"
)

effector tutorial

show(
    "afftfilt="
    "real='hypot(re,im) * sin(0)':"
    "imag='hypot(re,im) * cos(0)':"
    "win_size=512:"
    "overlap=0.75"
)

effector tutorial

show(
    "afftfilt="
    "real='hypot(re,im) * cos(2 * 3.14 * (random(0) * 2-1))':"
    "imag='hypot(re,im) * sin(2 * 3.14 * (random(1) * 2-1))':"
    "win_size=128:"
    "overlap=0.8"
)
# fmt: on

effector tutorial

颤音效果

https://ffmpeg.org/ffmpeg-filters.html#vibrato

show("vibrato=f=10:d=0.8")

effector tutorial

/pytorch/audio/ci_env/lib/python3.10/site-packages/IPython/lib/display.py:188: RuntimeWarning: invalid value encountered in cast
  return scaled.astype("<h").tobytes(), nchan

颤音效果

https://ffmpeg.org/ffmpeg-filters.html#tremolo

show("tremolo=f=8:d=0.8")

effector tutorial

晶体化器

https://ffmpeg.org/ffmpeg-filters.html#crystalizer

show("crystalizer")

effector tutorial

镶边效果

https://ffmpeg.org/ffmpeg-filters.html#flanger

show("flanger")

effector tutorial

相位器

https://ffmpeg.org/ffmpeg-filters.html#aphaser

show("aphaser")

effector tutorial

脉冲器

https://ffmpeg.org/ffmpeg-filters.html#apulsator

show("apulsator", stereo=True)

effector tutorial

haas

https://ffmpeg.org/ffmpeg-filters.html#haas

show("haas")

effector tutorial

编解码器

def show_multi(configs):
    results = []
    for config in configs:
        effector = AudioEffector(**config)
        results.append(effector.apply(waveform, int(sr)))

    num_configs = len(configs)
    figsize = (6.4, 0.3 + num_configs * 0.9)
    f, axes = plt.subplots(num_configs, 1, figsize=figsize, sharex=True)
    for result, ax in zip(results, axes):
        ax.specgram(result[:, 0], Fs=sr)
    f.set_tight_layout(True)

    return [Audio(r.numpy().T, rate=sr) for r in results]

ogg

results = show_multi(
    [
        {"format": "ogg"},
        {"format": "ogg", "encoder": "vorbis"},
        {"format": "ogg", "encoder": "opus"},
    ]
)

effector tutorial

ogg - 默认编码器 (flac)

results[0]

ogg - vorbis

results[1]

ogg - opus

results[2]

mp3

https://trac.ffmpeg.org/wiki/Encode/MP3

results = show_multi(
    [
        {"format": "mp3"},
        {"format": "mp3", "codec_config": CodecConfig(compression_level=1)},
        {"format": "mp3", "codec_config": CodecConfig(compression_level=9)},
        {"format": "mp3", "codec_config": CodecConfig(bit_rate=192_000)},
        {"format": "mp3", "codec_config": CodecConfig(bit_rate=8_000)},
        {"format": "mp3", "codec_config": CodecConfig(qscale=9)},
        {"format": "mp3", "codec_config": CodecConfig(qscale=1)},
    ]
)

effector tutorial