基于TensorFlow的歌曲曲风变换
背景：
在图像上的风格变换（Style Transfer,论文，基于Torch的代码实现：neural-style）同样可以应用于音频中。
本文采用的是英文TravelingLight.mp3和东风破.mp3作为音频源和参考源。考虑到内存有限，仅仅截取音频中的10s进行风格的变换。
代码：
# -*- coding: utf-8 -*-
__author__ = 'jason'
import tensorflow as tf
import librosa# 用来提取音频文件, 参看<中文语音识别>
import numpy as np
import os
import pdb
#import shlex # python2 pipes
# 音频文件路径
content_audio = "TravelingLight.mp3"
style_audio = "东风破.mp3"
# 为英文歌曲添加周杰伦风味
# 剪辑一段音频, 默认取开头的10s, 太大内存吃不消
def cut_audio(filename, start_pos='00:00:00', lens=10):
newfile = os.path.splitext(os.path.basename(filename))[0] + '_' + str(lens) + 's.mp3'
# 确保系统中已安装ffmpeg,这是ffmpeg的命令行方式,更加详细的使用方法可以google
cmd = "ffmpeg -i {} -ss {} -t {} -acodec copy {}".format(filename, start_pos, lens, newfile)
os.system(cmd)
return newfile
#上面的ffmpeg注意-acodec copy参数，否则会报错
content_audio_10s = cut_audio(content_audio, start_pos='00:00:33')
style_audio_10s = cut_audio(style_audio, start_pos='00:00:38')
#content_audio_10s = "TravelingLight_10s.mp3"
#style_audio_10s = "东风破_10s.mp3"
# Short Time Fourier Transform音频转spectrogram（把1维信号转为2维, 可以被视作图像）
# https://en.wikipedia.org/wiki/Short-time_Fourier_transform
N_FFT = 2048
def read_audio(filename):
x, fs = librosa.load(filename)
S = librosa.stft(x, N_FFT)
p = np.angle(S)
S = np.log1p(np.abs(S[:,:430]))
return S, fs
content_data, _ = read_audio(content_audio_10s)
style_data, fs = read_audio(style_audio_10s)
samples_n = content_data.shape[1] # 430
channels_n = style_data.shape[0] # 1025
style_data = style_data[:channels_n, :samples_n]
content_data_tf = np.ascontiguousarray(content_data.T[None,None,:,:])
style_data_tf = np.ascontiguousarray(style_data.T[None,None,:,:])
# filter shape "[filter_height, filter_width, in_channels, out_channels]"
N_FILTERS = 4096
std = np.sqrt(2) * np.sqrt(2.0 / ((channels_n + N_FILTERS) * 11))
kernel = np.random.randn(1, 11, channels_n, N_FILTERS)*std
# content and style features
g = tf.Graph()
with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
# data shape "[batch, in_height, in_width, in_channels]",
x = tf.placeholder('float32', [1, 1, samples_n, channels_n], name="x")
kernel_tf = tf.constant(kernel, name="kernel", dtype='float32')
conv = tf.nn.conv2d(x, kernel_tf, strides=[1, 1, 1, 1], padding="VALID", name="conv")
net = tf.nn.relu(conv)
content_features = net.eval(feed_dict={x: content_data_tf})
style_features = net.eval(feed_dict={x: style_data_tf})
features = np.reshape(style_features, (-1, N_FILTERS))
style_gram = np.matmul(features.T, features) / samples_n
# Optimize
ALPHA= 0.01 # ALPHA越大,content越占主导; 如果ALPHA为0,表示没有content
result = None
with tf.Graph().as_default():
learning_rate= 0.001
x = tf.Variable(np.random.randn(1, 1, samples_n, channels_n).astype(np.float32)*learning_rate, name="x")
kernel_tf = tf.constant(kernel, name="kernel", dtype='float32')
conv = tf.nn.conv2d(x, kernel_tf, strides=[1, 1, 1, 1], padding="VALID", name="conv")
net = tf.nn.relu(conv)
content_loss = ALPHA * 2 * tf.nn.l2_loss(net - content_features)
style_loss = 0
_, height, width, number = map(lambda i: i.value, net.get_shape())
size = height * width * number
feats = tf.reshape(net, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / samples_n
style_loss = 2 * tf.nn.l2_loss(gram - style_gram)
# loss
loss = content_loss + style_loss
opt = tf.contrib.opt.ScipyOptimizerInterface(loss, method='L-BFGS-B', options={'maxiter': 300})
# Optimization
init_op = tf.initialize_all_variables()#注意这里的初始化操作
with tf.Session() as sess:
#sess.run(tf.global_variables_initializer())#这是原来的初始化操作,会报错
sess.run(init_op)
opt.minimize(sess)
result = x.eval()
# 把spectrogram转回wav音频
audio = np.zeros_like(content_data)
audio[:channels_n,:] = np.exp(result[0,0].T) - 1
p = 2 * np.pi * np.random.random_sample(audio.shape) - np.pi
for i in range(500):
S = audio * np.exp(1j*p)
x = librosa.istft(S)
p = np.angle(librosa.stft(x, N_FFT))
librosa.output.write_wav("output.mp3", x, fs)结果：
运行后的数据结果下载