手把手教你语音识别（二）

127 阅读 0 评论 84 点赞

我是靠谱客的博主大气薯片，这篇文章主要介绍手把手教你语音识别（二），现在分享给大家，希望可以做个参考。

朋友们，之前的文章：手把手教你语音识别给大家讲过语音识别的流程，只不过，这些只是一个笼统的流程，下面开始从数据处理方面讲解怼入model之前都做了什么。

1、LogMelSpectrogramKaldi

第一步，就是读取后的数据要经过LogMel进行处理，这一部处理的目的一个是时域信号转频域信号，还有就是mel滤波器和log变换，具体这部分的至少可以参考：零基础入门语音识别: 一文详解MFCC特征（附python代码） - 知乎 (zhihu.com)

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-F9rAQUNz-1658977759410)(https://p3-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/3a2fb9dc132449ff8e709881ad6770e8~tplv-k3u1fbpfcp-watermark.image?)]

如上，可以看到，接之前的文章，数据读入的时候长度为58362，然后这个地方，先对读入的audio进行了一下变换，使用paddle.to_tensor增加了一个维度，然后下面就开始进行fbank提取特征了，咱们继续往下看。

2、Fbank

_get_waveform_and_window_properties

这里，是一个方法，用于得到窗口参数，可以看到，channel，也就是数据的维度，通道，这里默认给的是0，然后waveform这一行，相当于取了这个维度下面的所有值。然后可以看下图，window_shift的计算，也就是每次滑动的距离，咱们知道，之前默认是10ms，也就是时间10毫秒，但是他这个地方通过把16000Hz0.001变成了样本个数，频率时间=样本数，然后再乘以每秒的frame，实际上转化为滑动的样本数，window_size这也是一样的操作，经过这些处理后，padded_window_size 又把原来的400变成了512，这个是为何呢，其实是为了后期傅里叶变换更快，因为对于2的N次方，他更快。

复制代码

def _get_waveform_and_window_properties(
        waveform: Tensor,
        channel: int,
        sr: int,
        frame_shift: float,
        frame_length: float,
        round_to_power_of_two: bool,
        preemphasis_coefficient: float) -> Tuple[Tensor, int, int, int]:
    channel = max(channel, 0)
    assert channel < waveform.shape[0], (
        'Invalid channel {} for size {}'.format(channel, waveform.shape[0]))
    waveform = waveform[channel, :]  # size (n)
    window_shift = int(
        sr * frame_shift *
        0.001)  # pass frame_shift and frame_length in milliseconds
    window_size = int(sr * frame_length * 0.001)
    padded_window_size = _next_power_of_2(
        window_size) if round_to_power_of_two else window_size

assert 2 <= window_size <= len(waveform), (
        'choose a window size {} that is [2, {}]'.format(window_size,
                                                         len(waveform)))
    assert 0 < window_shift, '`window_shift` must be greater than 0'
    assert padded_window_size % 2 == 0, 'the padded `window_size` must be divisible by two.' 
                                        ' use `round_to_power_of_two` or change `frame_length`'
    assert 0. <= preemphasis_coefficient <= 1.0, '`preemphasis_coefficient` must be between [0,1]'
    assert sr > 0, '`sr` must be greater than zero'
    return waveform, window_shift, window_size, padded_window_size

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
def _get_waveform_and_window_properties(
        waveform: Tensor,
        channel: int,
        sr: int,
        frame_shift: float,
        frame_length: float,
        round_to_power_of_two: bool,
        preemphasis_coefficient: float) -> Tuple[Tensor, int, int, int]:
    channel = max(channel, 0)
    assert channel < waveform.shape[0], (
        'Invalid channel {} for size {}'.format(channel, waveform.shape[0]))
    waveform = waveform[channel, :]  # size (n)
    window_shift = int(
        sr * frame_shift *
        0.001)  # pass frame_shift and frame_length in milliseconds
    window_size = int(sr * frame_length * 0.001)
    padded_window_size = _next_power_of_2(
        window_size) if round_to_power_of_two else window_size

    assert 2 <= window_size <= len(waveform), (
        'choose a window size {} that is [2, {}]'.format(window_size,
                                                         len(waveform)))
    assert 0 < window_shift, '`window_shift` must be greater than 0'
    assert padded_window_size % 2 == 0, 'the padded `window_size` must be divisible by two.' 
                                        ' use `round_to_power_of_two` or change `frame_length`'
    assert 0. <= preemphasis_coefficient <= 1.0, '`preemphasis_coefficient` must be between [0,1]'
    assert sr > 0, '`sr` must be greater than zero'
    return waveform, window_shift, window_size, padded_window_size

在这里插入图片描述

_get_window

接着，一个函数是得到窗口的函数，把前面的输出作为输入，这个函数之前有一个函数叫_get_strided，这个函数的处理如下图，其中最核心的是第15行代码，m = 1 + (num_samples - window_size) // window_shift，这行其实就做了一件事，就是如果我用滑动窗口去截取每一段的frame，我能截取多少个，也就是滑动窗口取值的时候，能够取多少个值。

比如，如果num_samples == window_size，也就是我的窗口和你的样本一样长，那么就意味着我压根不用移动啊，因为一旦移动就越界了，所以此时1+0=1,也就是只有一个窗口值，frame=1，当然，如果窗口小，自然就按照上面去截取，如果不减去window_size的时候，截取的窗口是不重合的，为了保证数据冗余，存在重合，保证连续性，减去他再计算就对了，最终结果为363，这个数值一定要记住，后面会反复出现363。

在这里插入图片描述

复制代码

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
def _get_strided(waveform: Tensor,
                 window_size: int,
                 window_shift: int,
                 snip_edges: bool) -> Tensor:
    assert waveform.dim() == 1
    num_samples = waveform.shape[0]

    if snip_edges:
        if num_samples < window_size:
            return paddle.empty((0, 0), dtype=waveform.dtype)
        else:
            m = 1 + (num_samples - window_size) // window_shift
    else:
        reversed_waveform = paddle.flip(waveform, [0])
        m = (num_samples + (window_shift // 2)) // window_shift
        pad = window_size // 2 - window_shift // 2
        pad_right = reversed_waveform
        if pad > 0:
            pad_left = reversed_waveform[-pad:]
            waveform = paddle.concat((pad_left, waveform, pad_right), axis=0)
        else:
            waveform = paddle.concat((waveform[-pad:], pad_right), axis=0)

    return paddle.signal.frame(waveform, window_size, window_shift)[:, :m].T

复制代码

def _get_window(waveform: Tensor,
                padded_window_size: int,
                window_size: int,
                window_shift: int,
                window_type: str,
                blackman_coeff: float,
                snip_edges: bool,
                raw_energy: bool,
                energy_floor: float,
                dither: float,
                remove_dc_offset: bool,
                preemphasis_coefficient: float) -> Tuple[Tensor, Tensor]:
    dtype = waveform.dtype
    epsilon = _get_epsilon(dtype)

# (m, window_size)
    strided_input = _get_strided(waveform, window_size, window_shift,
                                 snip_edges)

if dither != 0.0:
        x = paddle.maximum(epsilon,
                           paddle.rand(strided_input.shape, dtype=dtype))
        rand_gauss = paddle.sqrt(-2 * x.log()) * paddle.cos(2 * math.pi * x)
        strided_input = strided_input + rand_gauss * dither

if remove_dc_offset:
        row_means = paddle.mean(strided_input, axis=1).unsqueeze(1)  # (m, 1)
        strided_input = strided_input - row_means

if raw_energy:
        signal_log_energy = _get_log_energy(strided_input, epsilon,
                                            energy_floor)  # (m)

if preemphasis_coefficient != 0.0:
        offset_strided_input = paddle.nn.functional.pad(
            strided_input.unsqueeze(0), (1, 0),
            data_format='NCL',
            mode='replicate').squeeze(0)  # (m, window_size + 1)
        strided_input = strided_input - preemphasis_coefficient * offset_strided_input[:, :
                                                                                       -1]

window_function = _feature_window_function(
        window_type, window_size, blackman_coeff,
        dtype).unsqueeze(0)  # (1, window_size)
    strided_input = strided_input * window_function  # (m, window_size)

# (m, padded_window_size)
    if padded_window_size != window_size:
        padding_right = padded_window_size - window_size
        strided_input = paddle.nn.functional.pad(
            strided_input.unsqueeze(0), (0, padding_right),
            data_format='NCL',
            mode='constant',
            value=0).squeeze(0)

if not raw_energy:
        signal_log_energy = _get_log_energy(strided_input, epsilon,
                                            energy_floor)  # size (m)

return strided_input, signal_log_energy

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
def _get_window(waveform: Tensor,
                padded_window_size: int,
                window_size: int,
                window_shift: int,
                window_type: str,
                blackman_coeff: float,
                snip_edges: bool,
                raw_energy: bool,
                energy_floor: float,
                dither: float,
                remove_dc_offset: bool,
                preemphasis_coefficient: float) -> Tuple[Tensor, Tensor]:
    dtype = waveform.dtype
    epsilon = _get_epsilon(dtype)

    # (m, window_size)
    strided_input = _get_strided(waveform, window_size, window_shift,
                                 snip_edges)

    if dither != 0.0:
        x = paddle.maximum(epsilon,
                           paddle.rand(strided_input.shape, dtype=dtype))
        rand_gauss = paddle.sqrt(-2 * x.log()) * paddle.cos(2 * math.pi * x)
        strided_input = strided_input + rand_gauss * dither

    if remove_dc_offset:
        row_means = paddle.mean(strided_input, axis=1).unsqueeze(1)  # (m, 1)
        strided_input = strided_input - row_means

    if raw_energy:
        signal_log_energy = _get_log_energy(strided_input, epsilon,
                                            energy_floor)  # (m)

    if preemphasis_coefficient != 0.0:
        offset_strided_input = paddle.nn.functional.pad(
            strided_input.unsqueeze(0), (1, 0),
            data_format='NCL',
            mode='replicate').squeeze(0)  # (m, window_size + 1)
        strided_input = strided_input - preemphasis_coefficient * offset_strided_input[:, :
                                                                                       -1]

    window_function = _feature_window_function(
        window_type, window_size, blackman_coeff,
        dtype).unsqueeze(0)  # (1, window_size)
    strided_input = strided_input * window_function  # (m, window_size)

    # (m, padded_window_size)
    if padded_window_size != window_size:
        padding_right = padded_window_size - window_size
        strided_input = paddle.nn.functional.pad(
            strided_input.unsqueeze(0), (0, padding_right),
            data_format='NCL',
            mode='constant',
            value=0).squeeze(0)

    if not raw_energy:
        signal_log_energy = _get_log_energy(strided_input, epsilon,
                                            energy_floor)  # size (m)

    return strided_input, signal_log_energy

在这里插入图片描述

如上，经过上面的一系列处理，strided_input变成了（363，512），363前面讲过了，是363个截取的窗口，512是每一个窗口的维度，也就是向量表达，前面本应该是400，变成512也讲了，目的是傅里叶变换计算更快。

rfft，开始傅里叶变换

在这里插入图片描述

这个实现，就不具体讲了，因为底层是C++实现的，咱们直接看结果。

在这里插入图片描述

结果是363，257，为啥维度正好减半，这个是rfft的操作导致的，具体可以搜索相关文章，总之现在已经完成变换，并且进行了平方，转为频谱了。

_get_mel_banks

这部分，就不详细讲了，其实就是对上面的数据再次处理，用滤波器处理，至于啥是滤波器，我给你举个例子，三棱镜知道不，咱们的太阳光是看不见，摸不着的白光知道不，光是七色混合的知道不，我现在想把白光里面的红光取出来做点别的事情，我咋办？对了，用三棱镜，这个三棱镜就是滤波器，把里面某个分量取出来，然后我可以用它作为特征，就这么个意思。