使用共振峰提取元音音素/从声音生成口型动画

视频效果

我前段时间研究了下从声音提取共振峰的方法。今天测试了下实际效果。

我使用一段33秒的女声视频，提取元音后使用静态视位图序列生成了一个视频，效果如下：
b站视频：
https://www.bilibili.com/video/BV1JD421H7m9/?vd_source=8abb7f0122649239c41b4c8acf458e47

https://www.bilibili.com/video/BV1i1421d7xU/?vd_source=8abb7f0122649239c41b4c8acf458e47

提取元音，生成视位序列

使用我前面一篇博客《使用python+librosa提取共振峰》中的方法可以提取音频文件的共振峰，使用三个共振峰作为三个坐标，和标准的元音共振峰求欧氏距离，距离最近的就是对应的元音。我使用了a\o\e\i\u五个元音。标准的元音共振峰数据我是参考的《实用语音识别基础—王炳锡》这本书的内容。
在这里插入图片描述
提取元音的代码如下：

import librosa
import numpy as np
from scipy.signal import lfilter, get_window
import matplotlib.pyplot as plt
import matplotlib.ticker as mtick
import math


def func_format(x, pos):
    return "%d" % (1000 * x)

class RhythmFeatures:
    """韵律学特征"""
    def __init__(self, input_file, sr=None, frame_len=512, n_fft=None, win_step=2 / 3, window="hamming"):
        """
        初始化
        :param input_file: 输入音频文件
        :param sr: 所输入音频文件的采样率，默认为None
        :param frame_len: 帧长，默认512个采样点(32ms,16kHz),与窗长相同
        :param n_fft: FFT窗口的长度，默认与窗长相同
        :param win_step: 窗移，默认移动2/3，512*2/3=341个采样点(21ms,16kHz)
        :param window: 窗类型，默认汉明窗
        """
        self.input_file = input_file
        self.frame_len = frame_len  # 帧长，单位采样点数
        self.wave_data, self.sr = librosa.load(self.input_file, sr=sr)
        self.window_len = frame_len  # 窗长512
        if n_fft is None:
            self.fft_num = self.window_len  # 设置NFFT点数与窗长相等
        else:
            self.fft_num = n_fft
        self.win_step = win_step
        self.hop_length = round(self.window_len * win_step)  # 重叠部分采样点数设置为窗长的1/3（1/3~1/2）,即帧移(窗移)2/3
        self.window = window

    def energy(self):
        """
        每帧内所有采样点的幅值平方和作为能量值
        :return: 每帧能量值，np.ndarray[shape=(1，n_frames), dtype=float64]
        """
        mag_spec = np.abs(librosa.stft(self.wave_data, n_fft=self.fft_num, hop_length=self.hop_length,
                                       win_length=self.frame_len, window=self.window))
        pow_spec = np.square(mag_spec)
        energy = np.sum(pow_spec, axis=0)
        energy = np.where(energy == 0, np.finfo(np.float64).eps, energy)  # 避免能量值为0，防止后续取log出错(eps是取非负的最小值)
        return energy

class Spectrogram:
    """声谱图（语谱图）特征"""
    def __init__(self, input_file, sr=None, frame_len=512, n_fft=None, win_step=2 / 3, window="hamming", preemph=0.97):
        """
        初始化
        :param input_file: 输入音频文件
        :param sr: 所输入音频文件的采样率，默认为None
        :param frame_len: 帧长，默认512个采样点(32ms,16kHz),与窗长相同
        :param n_fft: FFT窗口的长度，默认与窗长相同
        :param win_step: 窗移，默认移动2/3，512*2/3=341个采样点(21ms,16kHz)
        :param window: 窗类型，默认汉明窗
        :param preemph: 预加重系数,默认0.97
        """
        self.input_file = input_file
        self.wave_data, self.sr = librosa.load(self.input_file, sr=sr)  # 音频全部采样点的归一化数组形式数据
        self.wave_data = librosa.effects.preemphasis(self.wave_data, coef=preemph)  # 预加重，系数0.97
        self.window_len = frame_len  # 窗长512
        if n_fft is None:
            self.fft_num = self.window_len  # 设置NFFT点数与窗长相等
        else:
            self.fft_num = n_fft
        self.hop_length = round(self.window_len * win_step)  # 重叠部分采样点数设置为窗长的1/3（1/3~1/2）,即帧移(窗移)2/3
        self.window = window

    def get_magnitude_spectrogram(self):
        """
        获取幅值谱:fft后取绝对值
        :return: np.ndarray[shape=(1 + n_fft/2, n_frames), dtype=float32]，（257，全部采样点数/(512*2/3)+1）
        """
        # 频谱矩阵：行数=1 + n_fft/2=257，列数=帧数n_frames=全部采样点数/(512*2/3)+1（向上取整）
        # 快速傅里叶变化+汉明窗
        mag_spec = np.abs(librosa.stft(self.wave_data, n_fft=self.fft_num, hop_length=self.hop_length,
                                       win_length=self.window_len, window=self.window))
        return mag_spec


class QualityFeatures:
    """声音质量特征（音质）"""

    def __init__(self, input_file, sr=None, frame_len=512, n_fft=None, win_step=2 / 3, window="hamming"):
        """
        初始化
        :param input_file: 输入音频文件
        :param sr: 所输入音频文件的采样率，默认为None
        :param frame_len: 帧长，默认512个采样点(32ms,16kHz),与窗长相同
        :param n_fft: FFT窗口的长度，默认与窗长相同
        :param win_step: 窗移，默认移动2/3，512*2/3=341个采样点(21ms,16kHz)
        :param window: 窗类型，默认汉明窗
        """
        self.input_file = input_file
        self.frame_len = frame_len  # 帧长，单位采样点数
        self.wave_data, self.sr = librosa.load(self.input_file, sr=sr)
        print(f"wave_data.shape:{self.wave_data.shape}")
        self.n_fft = n_fft
        self.window_len = frame_len  # 窗长512
        self.win_step = win_step
        # 重叠部分采样点数设置为窗长的1/3（1/3~1/2）,即帧移(窗移)2/3
        self.hop_length = round(self.window_len * win_step)
        self.window = window

    def formant(self, ts_e=0.01, ts_f_d=200, ts_b_u=2000):
        """
        LPC求根法估计每帧前三个共振峰的中心频率及其带宽
        :param ts_e: 能量阈值：默认当能量超过0.01时认为可能会出现共振峰
        :param ts_f_d: 共振峰中心频率下阈值：默认当中心频率超过200，小于采样频率一半时认为可能会出现共振峰
        :param ts_b_u: 共振峰带宽上阈值：默认低于2000时认为可能会出现共振峰
        :return: F1/F2/F3、B1/B2/B3,每一列为一帧 F1/F2/F3或 B1/B2/B3，np.ndarray[shape=(3, n_frames), dtype=float64]
        """
        _data = lfilter([1., 0.83], [1], self.wave_data)  # 预加重0.83：高通滤波器
        inc_frame = self.hop_length  # 窗移
        n_frame = int(np.ceil(len(_data) / inc_frame))  # 分帧数
        n_pad = n_frame * self.window_len - len(_data)  # 末端补零数
        _data = np.append(_data, np.zeros(n_pad))  # 无法整除则末端补零
        win = get_window(self.window, self.window_len, fftbins=False)  # 获取窗函数
        formant_frq = []  # 所有帧组成的第1/2/3共振峰中心频率
        formant_bw = []  # 所有帧组成的第1/2/3共振峰带宽
        rym = RhythmFeatures(self.input_file, self.sr,
                             self.frame_len, self.n_fft, self.win_step, self.window)
        e = rym.energy()  # 获取每帧能量值
        e = e / np.max(e)  # 归一化
        for i in range(n_frame):
            f_i = _data[i * inc_frame:i * inc_frame + self.window_len]  # 分帧
            if np.all(f_i == 0):  # 避免上面的末端补零导致值全为0，防止后续求LPC线性预测误差系数出错(eps是取非负的最小值)
                f_i[0] = np.finfo(np.float64).eps
            f_i_win = f_i * win  # 加窗
            # 获取LPC线性预测误差系数，即滤波器分母多项式，阶数为 预期共振峰数3 *2+2，即想要得到F1-3
            a = librosa.lpc(f_i_win, order=8)
            rts = np.roots(a)  # 求LPC返回的预测多项式的根,为共轭复数对
            # 只保留共轭复数对一半，即虚数部分为+或-的根
            rts = np.array([r for r in rts if np.imag(r) >= 0])
            rts = np.where(rts == 0, np.finfo(np.float64).eps,
                           rts)  # 避免值为0，防止后续取log出错(eps是取非负的最小值)
            ang = np.arctan2(np.imag(rts), np.real(rts))  # 确定根对应的角(相位）
            # F(i) = ang(i)/(2*pi*T) = ang(i)*f/(2*pi)
            # 将以角度表示的rad/sample中的角频率转换为赫兹sample/s
            frq = ang * (self.sr / (2 * np.pi))
            indices = np.argsort(frq)  # 获取frq从小到大排序索引
            frequencies = frq[indices]  # frq从小到大排序
            # 共振峰的带宽由预测多项式零点到单位圆的距离表示: B(i) = -ln(r(i))/(pi*T) = -ln(abs(rts[i]))*f/pi
            bandwidths = -(self.sr / np.pi) * np.log(np.abs(rts[indices]))
            formant_f = []  # F1/F2/F3
            formant_b = []  # B1/B2/B3
            if e[i] > ts_e:  # 当能量超过ts_e时认为可能会出现共振峰
                # 采用共振峰频率大于ts_f_d小于self.sr/2赫兹，带宽小于ts_b_u赫兹的标准来确定共振峰
                for j in range(len(frequencies)):
                    if (ts_f_d < frequencies[j] < self.sr/2) and (bandwidths[j] < ts_b_u):
                        formant_f.append(frequencies[j])
                        formant_b.append(bandwidths[j])
                # 只取前三个共振峰
                if len(formant_f) < 3:  # 小于3个，则补nan
                    formant_f += ([np.nan] * (3 - len(formant_f)))
                else:  # 否则只取前三个
                    formant_f = formant_f[0:3]
                formant_frq.append(np.array(formant_f))  # 加入帧列表
                if len(formant_b) < 3:
                    formant_b += ([np.nan] * (3 - len(formant_b)))
                else:
                    formant_b = formant_b[0:3]
                formant_bw.append(np.array(formant_b))
            else:  # 能量过小，认为不会出现共振峰，此时赋值为nan
                formant_frq.append(np.array([np.nan, np.nan, np.nan]))
                formant_bw.append(np.array([np.nan, np.nan, np.nan]))
        formant_frq = np.array(formant_frq).T
        formant_bw = np.array(formant_bw).T
        # print(formant_frq.shape, np.nanmean(formant_frq, axis=1))
        # print(formant_bw.shape, np.nanmean(formant_bw, axis=1))
        return formant_frq, formant_bw

    def plot(self, show=True):
        """
        绘制语音波形曲线和log功率谱、共振峰叠加图
        :param show: 默认最后调用plt.show()，显示图形
        :return: None
        """
        plt.figure(figsize=(8, 6))
        # 以下绘制波形图
        plt.subplot(2, 1, 1)
        plt.title("Wave Form")
        plt.ylabel("Normalized Amplitude")
        plt.xticks([])
        audio_total_time = int(len(self.wave_data) / self.sr * 1000)  # 音频总时间ms
        plt.xlim(0, audio_total_time)
        plt.ylim(-1, 1)
        x = np.linspace(0, audio_total_time, len(self.wave_data))
        plt.plot(x, self.wave_data, c="b", lw=1)  # 语音波形曲线
        plt.axhline(y=0, c="pink", ls=":", lw=1)  # Y轴0线
        # 以下绘制灰度对数功率谱图
        plt.subplot(2, 1, 2)
        spec = Spectrogram(self.input_file, self.sr, self.frame_len,
                           self.n_fft, self.win_step, self.window, 0.83)
        log_power_spec = librosa.amplitude_to_db(
            spec.get_magnitude_spectrogram(), ref=np.max)
        librosa.display.specshow(log_power_spec[:, 1:], sr=self.sr, hop_length=self.hop_length,
                                 x_axis="s", y_axis="linear", cmap="gray_r")
        plt.title("Formants on Log-Power Spectrogram")
        plt.xlabel("Time/ms")
        plt.ylabel("Frequency/Hz")
        plt.gca().xaxis.set_major_formatter(mtick.FuncFormatter(func_format))
        # 以下在灰度对数功率谱图上叠加绘制共振峰点图
        formant_frq, __ = self.formant()  # 获取每帧共振峰中心频率
        color_p = {0: ".r", 1: ".y", 2: ".g"}  # 用不同颜色绘制F1-3点，对应红/黄/绿
        # X轴为对应的时间轴ms 从第0帧中间对应的时间开始，到总时长结束，间距为一帧时长
        x = np.linspace(0.5 * self.hop_length / self.sr,
                        audio_total_time / 1000, formant_frq.shape[1])
        for i in range(formant_frq.shape[0]):  # 依次绘制F1/F2/F3
            plt.plot(x, formant_frq[i, :], color_p[i], label="F" + str(i + 1))
        plt.legend(prop={'family': 'Times New Roman', 'size': 10}, loc="upper right",
                   framealpha=0.5, ncol=3, handletextpad=0.2, columnspacing=0.7)

        plt.tight_layout()
        if show:
            plt.show()




#男、女、儿童的标准元音共振峰位置，顺序：a\o\e\i\u
man_stand_f1 = [1000, 530, 540, 290, 380]
man_stand_f2 = [1160, 670, 1040, 2360, 440]
man_stand_f3 = [3120, 3310, 3170, 3570, 3660]
woman_stand_f1 = [1230, 720, 750, 320, 420]
woman_stand_f2 = [1350, 930, 1220, 2800, 650]
woman_stand_f3 = [2830, 2930, 3030, 3780, 3120]
child_stand_f1 = [1190, 850, 880, 390, 560]
child_stand_f2 = [1290, 1020, 1040, 3240, 810]
child_stand_f3 = [3650, 3580, 4100, 4260, 4340]

def Distance(f1, f2, f3, stand_f1, stand_f2, stand_f3):
    return math.sqrt(math.pow(f1-stand_f1, 2) + math.pow(f2-stand_f2, 2) + math.pow(f3-stand_f3, 2))  #欧式距离
    #return math.sqrt(math.pow(f1-stand_f1, 2) + math.pow(f2-stand_f2, 2))  #欧式距离

#根据和五个元音的共振峰的距离，返回对应的视位id， 1-5对应a\o\e\i\u
def Distance2Viseme(f1, f2, f3, stand_f1_vect, stand_f2_vect, stand_f3_vect):
    min_distance = 0
    min_index = 0
    dist_list = []
    for i in range(5):
        dist = Distance(f1, f2, f3, stand_f1_vect[i], stand_f2_vect[i], stand_f3_vect[i])
        dist_list.append(dist)
        if i == 0:
            min_distance = dist
            min_index = 0
        elif dist < min_distance:
            min_distance = dist
            min_index = i
    
    #print(f"f1:{f1}, f2:{f2}, f3:{f3}, min_index:{min_index}, min_distance:{min_distance}, dist_list:{dist_list}")

    return min_index+1, min_distance

# 将共振峰序列转化为视素序列
def Formants2Viseme(fmt_frq, voice_type="child"):
    stand_f1 = man_stand_f1
    stand_f2 = man_stand_f2
    stand_f3 = man_stand_f3
    if voice_type == "men":
        stand_f1 = man_stand_f1
        stand_f2 = man_stand_f2
        stand_f3 = man_stand_f3
    elif voice_type == "women":
        stand_f1 = woman_stand_f1
        stand_f2 = woman_stand_f2
        stand_f3 = woman_stand_f3
    elif voice_type == "child":
        stand_f1 = child_stand_f1
        stand_f2 = child_stand_f2
        stand_f3 = child_stand_f3

    viseme_list = []
    dist_list = []
    for index in range(fmt_frq.shape[1]): #遍历每一帧
        if np.isnan(fmt_frq[0][index]) or np.isnan(fmt_frq[1][index]) or np.isnan(fmt_frq[2][index]):
            viseme_list.append(0) #静止视位
            dist_list.append(-1)
        else:
            idv, dist = Distance2Viseme(fmt_frq[0][index], fmt_frq[1][index], fmt_frq[2][index], stand_f1, stand_f2, stand_f3)
            if 1:#dist < 1500:  #距离大于1000则认为是
                viseme_list.append(idv)
                dist_list.append(dist)
            else:
                viseme_list.append(0) #静止视位
                dist_list.append(-1)
    return viseme_list, dist_list

# 调用声音质量特征,获取共振峰
#quality_features = QualityFeatures("audios/audio_raw.wav")
quality_features = QualityFeatures("xxx_00004.wav", frame_len=600, win_step=2/3)  #窗口长度600， 步长400， 16K采样率下每秒40帧
#quality_features = QualityFeatures("audios/ae.wav")
fmt_frq, fmt_bw = quality_features.formant()  # 3个共振峰中心频率及其带宽

# 绘制波形图、功率谱，并叠加共振峰
#quality_features.plot(True)
print(f"fmt_frq.shape:{fmt_frq.shape}")
# print(f"fmt_frq:{fmt_frq}")

#将共振峰序列转换为视位序列。 视位0~5分别对应：sil\a\o\e\i\u
viseme_list, dist_list = Formants2Viseme(fmt_frq, voice_type="child")
print(f"viseme_list.len:{len(viseme_list)}")
print(f"dist_list.len:{len(dist_list)}")

print(f"viseme_list:{viseme_list}")
# print(f"dist_list:{dist_list}")

代码执行结果输出：

> python.exe .\test.py        
wave_data.shape:(535200,)
fmt_frq.shape:(3, 1338)
viseme_list.len:1338
dist_list.len:1338
viseme_list:[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 2, 4, 4, 4, 4, 4, 4, 2, 4, 0, 1, 4, 4, 4, 2, 4, 0, 0, 3, 3, 3, 5, 5, 3, 5, 0, 0, 0, 3, 3, 3, 3, 3, 4, 4, 4, 0, 4, 3, 2, 3, 3, 5, 3, 1, 2, 2, 1, 0, 0, 0, 0, 4, 4, 4, 3, 3, 3, 3, 3, 3, 3, 3, 3, 0, 4, 4, 2, 1, 1, 4, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 4, 2, 2, 4, 4, 4, 4, 5, 4, 4, 0, 4, 1, 2, 2, 4, 4, 4, 4, 4, 4, 1, 0, 0, 0, 3, 3, 2, 4, 4, 4, 4, 4, 2, 3, 2, 2, 3, 3, 3, 2, 2, 1, 1, 1, 4, 4, 0, 4, 0, 0, 0, 4, 4, 4, 0, 4, 0, 0, 2, 0, 4, 0, 0, 0, 0, 2, 2, 2, 2, 1, 1, 1, 1, 1, 4, 4, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 3, 4, 0, 0, 4, 4, 4, 2, 2, 2, 2, 3, 3, 3, 3, 3, 0, 0, 2, 2, 2, 2, 2, 4, 4, 4, 4, 4, 2, 4, 4, 0, 4, 1, 1, 5, 5, 0, 0, 2, 2, 2, 5, 2, 3, 3, 5, 4, 0, 4, 4, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 4, 4, 1, 4, 0, 4, 0, 2, 1, 3, 2, 0, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 4, 4, 4, 4, 4, 4, 4, 0, 0, 0, 0, 0, 4, 4, 4, 4, 5, 2, 2, 4, 4, 0, 0, 4, 4, 4, 1, 3, 3, 3, 2, 2, 3, 5, 3, 4, 0, 0, 0, 0, 3, 4, 0, 0, 0, 2, 1, 4, 0, 0, 0, 0, 0, 4, 4, 4, 4, 4, 4, 4, 3, 0, 5, 4, 4, 4, 0, 4, 0, 3, 3, 3, 5, 4, 4, 5, 5, 5, 5, 0, 4, 0, 4, 3, 1, 3, 3, 3, 5, 5, 4, 0, 0, 0, 3, 3, 3, 4, 0, 3, 3, 3, 0, 0, 4, 0, 4, 4, 3, 2, 3, 3, 3, 3, 3, 3, 0, 0, 0, 0, 4, 4, 3, 5, 5, 5, 0, 0, 4, 4, 4, 4, 4, 4, 4, 0, 3, 1, 3, 5, 2, 5, 5, 5, 3, 3, 3, 1, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 4, 1, 1, 1, 2, 4, 4, 0, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 3, 2, 2, 3, 3, 5, 4, 2, 2, 0, 0, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 0, 0, 0, 3, 3, 3, 3, 4, 4, 0, 0, 0, 0, 0, 4, 4, 4, 4, 4, 4, 4, 0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 0, 4, 4, 4, 4, 4, 4, 4, 4, 4, 2, 2, 0, 4, 0, 0, 0, 0, 3, 5, 5, 5, 5, 5, 0, 0, 5, 3, 3, 3, 3, 5, 3, 3, 3, 3, 3, 1, 3, 3, 0, 0, 4, 3, 3, 2, 2, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 4, 4, 1, 5, 4, 0, 0, 0, 3, 3, 3, 2, 2, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 4, 3, 5, 2, 2, 2, 1, 1, 1, 1, 1, 1, 2, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 3, 3, 3, 3, 3, 3, 3, 4, 0, 0, 4, 2, 1, 3, 3, 3, 3, 3, 5, 4, 0, 0, 0, 0, 5, 5, 0, 0, 0, 4, 2, 2, 2, 2, 4, 4, 0, 0, 0, 0, 4, 4, 4, 4, 4, 0, 0, 3, 3, 5, 5, 2, 5, 0, 0, 1, 1, 2, 3, 5, 2, 2, 2, 4, 0, 0, 0, 4, 4, 4, 4, 4, 0, 0, 4, 2, 2, 2, 4, 5, 0, 0, 0, 4, 2, 2, 2, 5, 5, 5, 5, 3, 0, 0, 0, 0, 4, 1, 3, 3, 3, 5, 0, 0, 0, 0, 5, 5, 3, 3, 4, 0, 0, 4, 1, 3, 3, 3, 2, 3, 4, 0, 4, 0, 4, 4, 3, 2, 2, 2, 3, 0, 0, 0, 0, 0, 3, 3, 5, 5, 3, 3, 5, 5, 4, 4, 0, 0, 0, 0, 4, 4, 4, 0, 3, 3, 3, 3, 3, 0, 4, 4, 4, 4, 4, 4, 1, 1, 1, 2, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 4, 4, 0, 0, 4, 4, 4, 4, 0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 0, 0, 5, 0, 3, 3, 2, 5, 5, 0, 0, 0, 5, 5, 5, 4, 0, 0, 0, 4, 4, 0, 0, 0, 3, 5, 3, 3, 3, 3, 5, 3, 3, 3, 3, 3, 2, 4, 0, 0, 0, 0, 0, 2, 2, 2, 3, 5, 5, 5, 2, 2, 0, 5, 0, 3, 3, 3, 3, 2, 0, 4, 0, 4, 0, 2, 2, 2, 3, 3, 3, 0, 0, 1, 2, 3, 3, 3, 4, 4, 4, 4, 4, 4, 0, 0, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 5, 0, 0, 5, 5, 3, 2, 2, 4, 0, 0, 0, 2, 2, 3, 3, 3, 2, 4, 0, 0, 4, 4, 4, 4, 4, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, 3, 3, 3, 3, 4, 4, 0, 0, 4, 4, 3, 5, 5, 5, 3, 4, 4, 4, 4, 0, 0, 3, 2, 3, 3, 5, 0, 0, 0, 0, 4, 3, 3, 3, 3, 3, 2, 3, 3, 2, 2, 4, 4, 0, 4, 4, 0, 4, 4, 1, 4, 4, 0, 0, 3, 5, 5, 0, 4, 0, 0, 3, 3, 3, 5, 5, 5, 0, 0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 1, 1, 4, 0, 1, 2, 2, 2, 1, 4, 4, 4, 4, 0, 4, 0, 3, 3, 3, 3, 3, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 3, 3, 5, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 4, 4, 1, 4, 4, 0, 1, 3, 2, 2, 3, 5, 4, 0, 0, 5, 3, 3, 3, 0, 0, 1, 4, 2, 2, 4, 4, 0, 0, 0, 4, 1, 1, 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 5, 0, 0, 3, 2, 2, 5, 3, 3, 3, 2, 2, 3, 3, 1, 1, 1, 3, 3, 4, 0, 0, 4, 4, 4, 4, 3, 3, 3, 3, 5, 4, 0, 0, 0, 3, 3, 5, 5, 5, 0, 0, 5, 2, 2, 2, 2, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

上面代码中检测共振峰的是这么一行：

quality_features = QualityFeatures("xxx_00004.wav", frame_len=600, win_step=2/3)  #窗口长度

由于音频采样率是16000，frame_len传入的是600，则音频分帧后每一个音频切片对应(600/16000)*1000=37.5ms。win_step为2/3，也就是音频帧中心间隔400次采样，对应25ms，所以每秒音频会产生40个语音切片，最终生成的视位fps对应40fps。

音频文件中共535200次采样，最终分帧数为1338。算出音频时长是1338x25ms=33450ms。

输出结果中最重要的就是viseme_list这个变量，它就是我们的视位序列。他的长度是1338，也就是共1338个视位。这个变量是下一步要用的。视位序列内容是0~5的数字，0表示静音状态，1-5分别表示a\o\e\i\u的嘴型视位。

代码中我同时列出了男性、女性和儿童的参考共振峰，发现我的这段测试音频用儿童的参考共振峰效果最好。

视位静止图片

从微软：将音素映射到视素可以看到很多因素对应的视位。我选取了a\o\e\i\u的对应的5张图片对应的视位，存到一个名称为viseme的目录下，如下
在这里插入图片描述
这里有个坑，那就是这些图片的像素尺寸不是严格一致的，需要在画图板中手工调整尺寸统一。我统一调整到(901, 859)。

将视位序列保存为视频

将上面执行输出的viseme_list复制过来，使用opencv将视位图片保存成视频。
由于上面生成的视位fps是40，我在生成视频前将视位id做了滤波，也就是每4个位置进行统计，压缩为一个视位，压缩后的视位就是4个位置上出现最多的视位。所以我生成的视频是10fps的。

从视位序列生成视位视频的代码如下

import cv2
import math

#将上一步的元音分析结果拷贝过来
# child, 3fs
viseme_list = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 2, 4, 4, 4, 4, 4, 4, 2, 4, 0, 1, 4, 4, 4, 2, 4, 0, 0, 3, 3, 3, 5, 5, 3, 5, 0, 0, 0, 3, 3, 3, 3, 3, 4, 4, 4, 0, 4, 3, 2, 3, 3, 5, 3, 1, 2, 2, 1, 0, 0, 0, 0, 4, 4, 4, 3, 3, 3, 3, 3, 3, 3, 3, 3, 0, 4, 4, 2, 1, 1, 4, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 4, 2, 2, 4, 4, 4, 4, 5, 4, 4, 0, 4, 1, 2, 2, 4, 4, 4, 4, 4, 4, 1, 0, 0, 0, 3, 3, 2, 4, 4, 4, 4, 4, 2, 3, 2, 2, 3, 3, 3, 2, 2, 1, 1, 1, 4, 4, 0, 4, 0, 0, 0, 4, 4, 4, 0, 4, 0, 0, 2, 0, 4, 0, 0, 0, 0, 2, 2, 2, 2, 1, 1, 1, 1, 1, 4, 4, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 3, 4, 0, 0, 4, 4, 4, 2, 2, 2, 2, 3, 3, 3, 3, 3, 0, 0, 2, 2, 2, 2, 2, 4, 4, 4, 4, 4, 2, 4, 4, 0, 4, 1, 1, 5, 5, 0, 0, 2, 2, 2, 5, 2, 3, 3, 5, 4, 0, 4, 4, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 4, 4, 1, 4, 0, 4, 0, 2, 1, 3, 2, 0, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 4, 4, 4, 4, 4, 4, 4, 0, 0, 0, 0, 0, 4, 4, 4, 4, 5, 2, 2, 4, 4, 0, 0, 4, 4, 4, 1, 3, 3, 3, 2, 2, 3, 5, 3, 4, 0, 0, 0, 0, 3, 4, 0, 0, 0, 2, 1, 4, 0, 0, 0, 0, 0, 4, 4, 4, 4, 4, 4, 4, 3, 0, 5, 4, 4, 4, 0, 4, 0, 3, 3, 3, 5, 4, 4, 5, 5, 5, 5, 0, 4, 0, 4, 3, 1, 3, 3, 3, 5, 5, 4, 0, 0, 0, 3, 3, 3, 4, 0, 3, 3, 3, 0, 0, 4, 0, 4, 4, 3, 2, 3, 3, 3, 3, 3, 3, 0, 0, 0, 0, 4, 4, 3, 5, 5, 5, 0, 0, 4, 4, 4, 4, 4, 4, 4, 0, 3, 1, 3, 5, 2, 5, 5, 5, 3, 3, 3, 1, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 4, 1, 1, 1, 2, 4, 4, 0, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 3, 2, 2, 3, 3, 5, 4, 2, 2, 0, 0, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 0, 0, 0, 3, 3, 3, 3, 4, 4, 0, 0, 0, 0, 0, 4, 4, 4, 4, 4, 4, 4, 0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 0, 4, 4, 4, 4, 4, 4, 4, 4, 4, 2, 2, 0, 4, 0, 0, 0, 0, 3, 5, 5, 5, 5, 5, 0, 0, 5, 3, 3, 3, 3, 5, 3, 3, 3, 3, 3, 1, 3, 3, 0, 0, 4, 3, 3, 2, 2, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 4, 4, 1, 5, 4, 0, 0, 0, 3, 3, 3, 2, 2, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 4, 3, 5, 2, 2, 2, 1, 1, 1, 1, 1, 1, 2, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 3, 3, 3, 3, 3, 3, 3, 4, 0, 0, 4, 2, 1, 3, 3, 3, 3, 3, 5, 4, 0, 0, 0, 0, 5, 5, 0, 0, 0, 4, 2, 2, 2, 2, 4, 4, 0, 0, 0, 0, 4, 4, 4, 4, 4, 0, 0, 3, 3, 5, 5, 2, 5, 0, 0, 1, 1, 2, 3, 5, 2, 2, 2, 4, 0, 0, 0, 4, 4, 4, 4, 4, 0, 0, 4, 2, 2, 2, 4, 5, 0, 0, 0, 4, 2, 2, 2, 5, 5, 5, 5, 3, 0, 0, 0, 0, 4, 1, 3, 3, 3, 5, 0, 0, 0, 0, 5, 5, 3, 3, 4, 0, 0, 4, 1, 3, 3, 3, 2, 3, 4, 0, 4, 0, 4, 4, 3, 2, 2, 2, 3, 0, 0, 0, 0, 0, 3, 3, 5, 5, 3, 3, 5, 5, 4, 4, 0, 0, 0, 0, 4, 4, 4, 0, 3, 3, 3, 3, 3, 0, 4, 4, 4, 4, 4, 4, 1, 1, 1, 2, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 4, 4, 0, 0, 4, 4, 4, 4, 0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 0, 0, 5, 0, 3, 3, 2, 5, 5, 0, 0, 0, 5, 5, 5, 4, 0, 0, 0, 4, 4, 0, 0, 0, 3, 5, 3, 3, 3, 3, 5, 3, 3, 3, 3, 3, 2, 4, 0, 0, 0, 0, 0, 2, 2, 2, 3, 5, 5, 5, 2, 2, 0, 5, 0, 3, 3, 3, 3, 2, 0, 4, 0, 4, 0, 2, 2, 2, 3, 3, 3, 0, 0, 1, 2, 3, 3, 3, 4, 4, 4, 4, 4, 4, 0, 0, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 5, 0, 0, 5, 5, 3, 2, 2, 4, 0, 0, 0, 2, 2, 3, 3, 3, 2, 4, 0, 0, 4, 4, 4, 4, 4, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, 3, 3, 3, 3, 4, 4, 0, 0, 4, 4, 3, 5, 5, 5, 3, 4, 4, 4, 4, 0, 0, 3, 2, 3, 3, 5, 0, 0, 0, 0, 4, 3, 3, 3, 3, 3, 2, 3, 3, 2, 2, 4, 4, 0, 4, 4, 0, 4, 4, 1, 4, 4, 0, 0, 3, 5, 5, 0, 4, 0, 0, 3, 3, 3, 5, 5, 5, 0, 0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 1, 1, 4, 0, 1, 2, 2, 2, 1, 4, 4, 4, 4, 0, 4, 0, 3, 3, 3, 3, 3, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 3, 3, 5, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 4, 4, 1, 4, 4, 0, 1, 3, 2, 2, 3, 5, 4, 0, 0, 5, 3, 3, 3, 0, 0, 1, 4, 2, 2, 4, 4, 0, 0, 0, 4, 1, 1, 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 5, 0, 0, 3, 2, 2, 5, 3, 3, 3, 2, 2, 3, 3, 1, 1, 1, 3, 3, 4, 0, 0, 4, 4, 4, 4, 3, 3, 3, 3, 5, 4, 0, 0, 0, 3, 3, 5, 5, 5, 0, 0, 5, 2, 2, 2, 2, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]


# 视位帧间隔，单位ms.  25对应fps是40
viseme_time = 25
# 对视位进行滤波的时候，会对每filter_interval个进行投票，得票多的视位为当前的视位
filter_interval = 4

#读取视位的图片
viseme_sil_mat = cv2.imread("./viseme/sil.jpg")
viseme_a_mat = cv2.imread("./viseme/a.jpg")
viseme_o_mat = cv2.imread("./viseme/o.jpg")
viseme_e_mat = cv2.imread("./viseme/e.jpg")
viseme_i_mat = cv2.imread("./viseme/i.jpg")
viseme_u_mat = cv2.imread("./viseme/u.jpg")

viseme_mat = [viseme_sil_mat, viseme_a_mat, viseme_o_mat, viseme_e_mat, viseme_i_mat, viseme_u_mat]

#这里实现滤波，在filter_interval个帧中选投票个数最多的视位作为滤波后的新视位
#假设输入视位是N个，滤波后的视位是N/filter_interval个
def filter4viseme(old_viseme_list, filter_interval=5):
    new_viseme_list = []
    
    new_frame_num = math.ceil(len(old_viseme_list)/filter_interval) # 向上取整
    for i in range(new_frame_num):
        vote_list = [0, 0, 0, 0, 0, 0]  #6个视位的投票箱
        for x in range(i*filter_interval, (i+1)*filter_interval):
            if x >= len(old_viseme_list):  #如果超出范围则停止循环
                break
            viseme_id = old_viseme_list[x]
            vote_list[viseme_id] += 1

        max_vote = max(vote_list) #求列表最大值
        max_viseme_id = vote_list.index(max_vote) #求最大值对应的下标
        print(f"max_viseme_id:{max_viseme_id}, max_vote:{max_vote}, vote_list:{vote_list}")

        new_viseme_list.append(max_viseme_id)

    print(f"len(new_viseme_list):{len(new_viseme_list)}")
    print(f"new_viseme_list:{new_viseme_list}")

    return new_viseme_list

# 对视位进行滤波平滑，每filter_interval压缩为1个
filted_viseme_list = filter4viseme(viseme_list, filter_interval)

fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter('testwrite-child10fps3f.mp4', fourcc, 1000/viseme_time/filter_interval, (901, 859), True)
# out = cv2.VideoWriter('testwrite-new.avi', fourcc, 20, (901, 859), True)

#读取压缩后的视位id，按照顺序写入视频文件
for index in range(len(filted_viseme_list)):
    viseme_id = filted_viseme_list[index]
    frame = viseme_mat[viseme_id]

    out.write(frame)
    cv2.imshow("viseme", frame)
    print(f"{index}/{len(filted_viseme_list)}, viseme_id={viseme_id} | ")

    if cv2.waitKey(1) == 27:
        break

out.release()
cv2.destroyAllWindows()