Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
DeepSpeech
提交
4eb780ad
D
DeepSpeech
项目概览
PaddlePaddle
/
DeepSpeech
12 个月 前同步成功
通知
205
Star
8425
Fork
1598
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
245
列表
看板
标记
里程碑
合并请求
3
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
D
DeepSpeech
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
245
Issue
245
列表
看板
标记
里程碑
合并请求
3
合并请求
3
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
4eb780ad
编写于
3月 04, 2022
作者:
K
KP
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add reference and doc string.
上级
5201c59f
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
94 addition
and
36 deletion
+94
-36
paddleaudio/paddleaudio/compliance/kaldi.py
paddleaudio/paddleaudio/compliance/kaldi.py
+1
-0
paddleaudio/paddleaudio/features/layers.py
paddleaudio/paddleaudio/features/layers.py
+86
-30
paddleaudio/paddleaudio/functional/functional.py
paddleaudio/paddleaudio/functional/functional.py
+7
-6
未找到文件。
paddleaudio/paddleaudio/compliance/kaldi.py
浏览文件 @
4eb780ad
...
...
@@ -11,6 +11,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified from torchaudio(https://github.com/pytorch/audio)
import
math
from
typing
import
Tuple
...
...
paddleaudio/paddleaudio/features/layers.py
浏览文件 @
4eb780ad
...
...
@@ -44,22 +44,22 @@ class Spectrogram(nn.Layer):
The spectorgram is defined as the complex norm of the short-time
Fourier transformation.
Parameters:
n_fft(int): the number of frequency components of the discrete Fourier transform.
n_fft
(int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
hop_length
(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window(str): the name of the window function applied to the single before the Fourier transform.
window
(str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
center
(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
pad_mode
(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'. The default value is 'reflect'.
dtype(str): the data type of input and window.
dtype
(str): the data type of input and window.
Notes:
The Spectrogram transform relies on STFT transform to compute the spectrogram.
By default, the weights are not learnable. To fine-tune the Fourier coefficients,
...
...
@@ -190,39 +190,39 @@ class LogMelSpectrogram(nn.Layer):
"""Compute log-mel-spectrogram(also known as LogFBank) feature of a given signal,
typically an audio waveform.
Parameters:
sr(int): the audio sample rate.
sr
(int): the audio sample rate.
The default value is 22050.
n_fft(int): the number of frequency components of the discrete Fourier transform.
n_fft
(int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
hop_length
(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window(str): the name of the window function applied to the single before the Fourier transform.
window
(str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
center
(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
pad_mode
(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'.
The default value is 'reflect'.
n_mels(int): the mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zeros.
ref_value(float): the reference value. If smaller than 1.0, the db level
htk(bool): whether to use HTK formula in computing fbank matrix.
norm(str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
n_mels (int): the mel bins.
f_min (float): the lower cut-off frequency, below which the filter response is zero.
f_max (float): the upper cut-off frequency, above which the filter response is zeros.
htk (bool): whether to use HTK formula in computing fbank matrix.
norm (str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
You can specify norm=1.0/2.0 to use customized p-norm normalization.
dtype(str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
amin(float): the minimum value of input magnitude, below which the input of the signal will be pulled up accordingly.
ref_value (float): the reference value. If smaller than 1.0, the db level
amin (float): the minimum value of input magnitude, below which the input of the signal will be pulled up accordingly.
Otherwise, the db level is pushed down.
magnitude is clipped(to amin). For numerical stability, set amin to a larger value,
e.g., 1e-3.
top_db(float): the maximum db value of resulting spectrum, above which the
top_db
(float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
dtype (str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
"""
super
(
LogMelSpectrogram
,
self
).
__init__
()
...
...
@@ -260,24 +260,80 @@ class MFCC(nn.Layer):
def
__init__
(
self
,
sr
:
int
=
22050
,
n_mfcc
:
int
=
40
,
norm
:
str
=
'ortho'
,
dtype
:
str
=
paddle
.
float32
,
**
kwargs
):
n_fft
:
int
=
512
,
hop_length
:
Optional
[
int
]
=
None
,
win_length
:
Optional
[
int
]
=
None
,
window
:
str
=
'hann'
,
center
:
bool
=
True
,
pad_mode
:
str
=
'reflect'
,
n_mels
:
int
=
64
,
f_min
:
float
=
50.0
,
f_max
:
Optional
[
float
]
=
None
,
htk
:
bool
=
False
,
norm
:
Union
[
str
,
float
]
=
'slaney'
,
ref_value
:
float
=
1.0
,
amin
:
float
=
1e-10
,
top_db
:
Optional
[
float
]
=
None
,
dtype
:
str
=
paddle
.
float32
):
"""Compute mel frequency cepstral coefficients(MFCCs) feature of given waveforms.
Parameters:
sr(int): the audio sample rate.
The default value is 22050.
n_mfcc (int, optional): Number of cepstra in MFCC. Defaults to 40.
norm(str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
n_fft (int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length (int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window (str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center (bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode (str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'.
The default value is 'reflect'.
n_mels (int): the mel bins.
f_min (float): the lower cut-off frequency, below which the filter response is zero.
f_max (float): the upper cut-off frequency, above which the filter response is zeros.
htk (bool): whether to use HTK formula in computing fbank matrix.
norm (str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
You can specify norm=1.0/2.0 to use customized p-norm normalization.
dtype(str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
ref_value (float): the reference value. If smaller than 1.0, the db level
amin (float): the minimum value of input magnitude, below which the input of the signal will be pulled up accordingly.
Otherwise, the db level is pushed down.
magnitude is clipped(to amin). For numerical stability, set amin to a larger value,
e.g., 1e-3.
top_db (float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
dtype (str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
"""
super
(
MFCC
,
self
).
__init__
()
self
.
_log_melspectrogram
=
LogMelSpectrogram
(
sr
=
sr
,
**
kwargs
)
self
.
dct_matrix
=
create_dct
(
n_mfcc
=
n_mfcc
,
n_mels
=
self
.
_log_melspectrogram
.
n_mels
,
norm
=
norm
)
assert
n_mfcc
<=
n_mels
,
'n_mfcc cannot be larger than n_mels: %d vs %d'
%
(
n_mfcc
,
n_mels
)
self
.
_log_melspectrogram
=
LogMelSpectrogram
(
sr
=
sr
,
n_fft
=
n_fft
,
hop_length
=
hop_length
,
win_length
=
win_length
,
window
=
window
,
center
=
center
,
pad_mode
=
pad_mode
,
n_mels
=
n_mels
,
f_min
=
f_min
,
f_max
=
f_max
,
htk
=
htk
,
norm
=
norm
,
ref_value
=
ref_value
,
amin
=
amin
,
top_db
=
top_db
,
dtype
=
dtype
)
self
.
dct_matrix
=
create_dct
(
n_mfcc
=
n_mfcc
,
n_mels
=
n_mels
,
dtype
=
dtype
)
self
.
register_buffer
(
'dct_matrix'
,
self
.
dct_matrix
)
def
forward
(
self
,
x
):
...
...
paddleaudio/paddleaudio/functional/functional.py
浏览文件 @
4eb780ad
...
...
@@ -242,14 +242,15 @@ def power_to_db(magnitude: paddle.Tensor,
def
create_dct
(
n_mfcc
:
int
,
n_mels
:
int
,
norm
:
Optional
[
str
]
=
'ortho'
,
dtype
:
Optional
[
str
]
=
paddle
.
float32
):
"""[summary]
dtype
:
Optional
[
str
]
=
paddle
.
float32
)
->
paddle
.
Tensor
:
"""Create a discrete cosine transform(DCT) matrix.
Parameters:
n_mfcc (int):
[description]
n_mels (int):
[description]
norm (str, optional):
[description]
. Defaults to 'ortho'.
n_mfcc (int):
Number of mel frequency cepstral coefficients.
n_mels (int):
Number of mel filterbanks.
norm (str, optional):
Normalizaiton type
. Defaults to 'ortho'.
Returns:
[type]: [description]
Tensor: The DCT matrix with shape (n_mels, n_mfcc).
"""
n
=
paddle
.
arange
(
n_mels
,
dtype
=
dtype
)
k
=
paddle
.
arange
(
n_mfcc
,
dtype
=
dtype
).
unsqueeze
(
1
)
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录