python 声质変换する

Posted 2021-05-10

#coding: utf-8
import os
import sys
import subprocess
import struct
import numpy as np
from scipy.stats import multivariate_normal
from sklearn.externals import joblib

# メルケプストラム次数
# 実際はパワー項を追加して26次元ベクトルになる
m = 25

# GMMのコンポーネント数
K = 32

def extract_mcep(wav_file, mcep_file, ascii=False):
    if ascii:
        cmd = "bcut +s -s 22 %s | x2x +sf | frame -p 80 | window | mcep -m %d -a 0.42 | x2x +fa%d > %s" % (wav_file, m, m + 1, mcep_file)
    else:
        cmd = "bcut +s -s 22 %s | x2x +sf | frame -p 80 | window | mcep -m %d -a 0.42 > %s" % (wav_file, m, mcep_file)
    subprocess.call(cmd, shell=True)

def extract_pitch(wav_file, pitch_file):
    cmd = "bcut +s -s 22 %s | x2x +sf | pitch -a 1 -s 16 -p 80 > %s" % (wav_file, pitch_file)
    subprocess.call(cmd, shell=True)

def convert_mcep(source_mcep_file, converted_mcep_file, gmm):
    source_mcep = np.loadtxt(source_mcep_file)
    fp = open(converted_mcep_file, "wb")

    d = m + 1

    # 式9の多次元正規分布のオブジェクトを作成しておく
    gauss = []
    for k in range(K):
        gauss.append(multivariate_normal(gmm.means_[k, 0:d], gmm.covars_[k, 0:d, 0:d]))

    # 式11のフレームtに依存しない項を計算しておく
    ss = []
    for k in range(K):
        ss.append(np.dot(gmm.covars_[k, d:, 0:d], np.linalg.inv(gmm.covars_[k, 0:d, 0:d])))

    # 各フレームをGMMで変形する
    for t in range(len(source_mcep)):
        x_t = source_mcep[t]
        y_t = convert_frame(x_t, gmm, gauss, ss)
        fp.write(struct.pack('f' * (m+1), *y_t))
    fp.close()

# 式(13)の計算
def convert_frame(x, gmm, gauss, ss):
    # 式(9)の分母だけ先に計算
    denom = np.zeros(K)
    for n in range(K):
        denom[n] = gmm.weights_[n] * gauss[n].pdf(x)

    y = np.zeros_like(x)
    for k in range(K):
        y += P(k, x, gmm, gauss, denom) * E(k, x, gmm, ss)
    return y

# 式(9)の計算
def P(k, x, gmm, gauss, denom):
    return denom[k] / np.sum(denom)

# 式(11)の計算
def E(k, x, gmm, ss):
    d = m + 1
    return gmm.means_[k, d:] + np.dot(ss[k], x - gmm.means_[k, 0:d])

def synthesis(pitch_file, mcep_file, wav_file):
    cmd = "excite -p 80 %s | mlsadf -m %d -a 0.42 -p 80 %s | clip -y -32000 32000 | x2x +fs > temp.raw" % (pitch_file, m, mcep_file)
    print cmd
    subprocess.call(cmd, shell=True)

    cmd = "sox -e signed-integer -c 1 -b 16 -r 16000 temp.raw %s" % (wav_file)
    print cmd
    subprocess.call(cmd, shell=True)

    os.remove("temp.raw")

if __name__ == "__main__":
    if len(sys.argv) != 4:
        print "usage: python train_gmm.py <source_wav_file> <gmm_file> <converted_wav_file>"
        exit()

    source_wav_file = sys.argv[1]
    gmm_file = sys.argv[2]
    converted_wav_file = sys.argv[3]

    # 変換元のwavファイルからメルケプストラムとピッチを抽出
    # numpyで読みやすいようにアスキー形式で保存
    print "extract mcep ..."
    source_mcep_file = "source.mcep_ascii"
    extract_mcep(source_wav_file, source_mcep_file, ascii=True)

    print "extract pitch ..."
    source_pitch_file = "source.pitch"
    extract_pitch(source_wav_file, source_pitch_file)

    # GMMをロード
    gmm = joblib.load(gmm_file)

    # 変換元のメルケプストラムをGMMで変換
    # SPTKで合成できるようにバイナリ形式で保存
    print "convert mcep ..."
    converted_mcep_file = "converted.mcep"
    convert_mcep(source_mcep_file, converted_mcep_file, gmm)

    # 変換元のピッチと変換したメルケプストラムから再合成
    print "synthesis ..."
    synthesis(source_pitch_file, converted_mcep_file, converted_wav_file)

    # 一時ファイルを削除
    os.remove(source_mcep_file)
    os.remove(source_pitch_file)
    os.remove(converted_mcep_file)