Merge branch 'develop' into develop

pull/1646/head
Honei 3 years ago committed by GitHub
commit 83310b6379
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

@ -180,7 +180,7 @@ Via the easy-to-use, efficient, flexible and scalable implementation, our vision
2021.12.14: We would like to have an online courses to introduce basics and research of speech, as well as code practice with `paddlespeech`. Please pay attention to our [Calendar](https://www.paddlepaddle.org.cn/live).
--->
- 👏🏻 2022.03.28: PaddleSpeech Server is available for Audio Classification, Automatic Speech Recognition and Text-to-Speech.
- 👏🏻 2022.03.28: PaddleSpeech CLI is available for Speaker Verfication.
- 👏🏻 2022.03.28: PaddleSpeech CLI is available for Speaker Verification.
- 🤗 2021.12.14: Our PaddleSpeech [ASR](https://huggingface.co/spaces/KPatrick/PaddleSpeechASR) and [TTS](https://huggingface.co/spaces/KPatrick/PaddleSpeechTTS) Demos on Hugging Face Spaces are available!
- 👏🏻 2021.12.10: PaddleSpeech CLI is available for Audio Classification, Automatic Speech Recognition, Speech Translation (English to Chinese) and Text-to-Speech.

@ -80,7 +80,7 @@ PANN | ESC-50 |[pann-esc50](../../examples/esc50/cls0)|[esc50_cnn6.tar.gz](https
Model Type | Dataset| Example Link | Pretrained Models | Static Models
:-------------:| :------------:| :-----: | :-----: | :-----:
PANN | VoxCeleb| [voxceleb_ecapatdnn](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/voxceleb/sv0) | [ecapatdnn.tar.gz](https://paddlespeech.bj.bcebos.com/vector/voxceleb/sv0_ecapa_tdnn_voxceleb12_ckpt_0_1_1.tar.gz) | -
PANN | VoxCeleb| [voxceleb_ecapatdnn](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/voxceleb/sv0) | [ecapatdnn.tar.gz](https://paddlespeech.bj.bcebos.com/vector/voxceleb/sv0_ecapa_tdnn_voxceleb12_ckpt_0_2_0.tar.gz) | -
## Punctuation Restoration Models
Model Type | Dataset| Example Link | Pretrained Models

@ -4,7 +4,7 @@
对于声音分类任务传统机器学习的一个常用做法是首先人工提取音频的时域和频域的多种特征并做特征选择、组合、变换等然后基于SVM或决策树进行分类。而端到端的深度学习则通常利用深度网络如RNNCNN等直接对声间波形(waveform)或时频特征(time-frequency)进行特征学习(representation learning)和分类预测。
在IEEE ICASSP 2017 大会上,谷歌开放了一个大规模的音频数据集[Audioset](https://research.google.com/audioset/)。该数据集包含了 632 类的音频类别以及 2,084,320 条人工标记的每段 10 秒长度的声音剪辑片段来源于YouTube视频。目前该数据集已经有210万个已标注的视频数据5800小时的音频数据经过标记的声音样本的标签类别为527。
在IEEE ICASSP 2017 大会上,谷歌开放了一个大规模的音频数据集[Audioset](https://research.google.com/audioset/)。该数据集包含了 632 类的音频类别以及 2,084,320 条人工标记的每段 **10 秒**长度的声音剪辑片段来源于YouTube视频。目前该数据集已经有 210万 个已标注的视频数据5800 小时的音频数据,经过标记的声音样本的标签类别为 527。
`PANNs`([PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition](https://arxiv.org/pdf/1912.10211.pdf))是基于Audioset数据集训练的声音分类/识别的模型。经过预训练后模型可以用于提取音频的embbedding。本示例将使用`PANNs`的预训练模型Finetune完成声音分类的任务。
@ -19,7 +19,7 @@ PaddleAudio提供了PANNs的CNN14、CNN10和CNN6的预训练模型可供用
## 数据集
[ESC-50: Dataset for Environmental Sound Classification](https://github.com/karolpiczak/ESC-50) 是一个包含有 2000 个带标签的环境声音样本,音频样本采样率为 44,100Hz 的单通道音频文件,所有样本根据标签被划分为 50 个类别,每个类别有 40 个样本。
[ESC-50: Dataset for Environmental Sound Classification](https://github.com/karolpiczak/ESC-50) 是一个包含有 2000 个带标签的时长为 **5 秒**环境声音样本,音频样本采样率为 44,100Hz 的单通道音频文件,所有样本根据标签被划分为 50 个类别,每个类别有 40 个样本。
## 模型指标

@ -4,4 +4,4 @@
| Model | Number of Params | Release | Config | dim | Test set | Cosine | Cosine + S-Norm |
| --- | --- | --- | --- | --- | --- | --- | ---- |
| ECAPA-TDNN | 85M | 0.1.2 | conf/ecapa_tdnn.yaml |192 | test | 1.02 | 0.95 |
| ECAPA-TDNN | 85M | 0.2.0 | conf/ecapa_tdnn.yaml |192 | test | 1.02 | 0.95 |

@ -14,4 +14,3 @@
from .dtw import dtw_distance
from .eer import compute_eer
from .eer import compute_minDCF
from .mcd import mcd_distance

@ -1,63 +0,0 @@
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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.
from typing import Callable
import mcd.metrics_fast as mt
import numpy as np
from mcd import dtw
__all__ = [
'mcd_distance',
]
def mcd_distance(xs: np.ndarray,
ys: np.ndarray,
cost_fn: Callable=mt.logSpecDbDist) -> float:
"""Mel cepstral distortion (MCD), dtw distance.
Dynamic Time Warping.
Uses dynamic programming to compute:
Examples:
.. code-block:: python
wps[i, j] = cost_fn(xs[i], ys[j]) + min(
wps[i-1, j ], // vertical / insertion / expansion
wps[i , j-1], // horizontal / deletion / compression
wps[i-1, j-1]) // diagonal / match
dtw = sqrt(wps[-1, -1])
Cost Function:
Examples:
.. code-block:: python
logSpecDbConst = 10.0 / math.log(10.0) * math.sqrt(2.0)
def logSpecDbDist(x, y):
diff = x - y
return logSpecDbConst * math.sqrt(np.inner(diff, diff))
Args:
xs (np.ndarray): ref sequence, [T,D]
ys (np.ndarray): hyp sequence, [T,D]
cost_fn (Callable, optional): Cost function. Defaults to mt.logSpecDbDist.
Returns:
float: dtw distance
"""
min_cost, path = dtw.dtw(xs, ys, cost_fn)
return min_cost

@ -19,7 +19,7 @@ from setuptools.command.install import install
from setuptools.command.test import test
# set the version here
VERSION = '0.2.0'
VERSION = '0.2.1'
# Inspired by the example at https://pytest.org/latest/goodpractises.html
@ -83,8 +83,7 @@ setuptools.setup(
python_requires='>=3.6',
install_requires=[
'numpy >= 1.15.0', 'scipy >= 1.0.0', 'resampy >= 0.2.2',
'soundfile >= 0.9.0', 'colorlog', 'dtaidistance == 2.3.1', 'mcd >= 0.4',
'pathos'
'soundfile >= 0.9.0', 'colorlog', 'dtaidistance == 2.3.1', 'pathos'
],
extras_require={
'test': [

@ -43,7 +43,7 @@ pretrained_models = {
# "paddlespeech vector --task spk --model ecapatdnn_voxceleb12-16k --sr 16000 --input ./input.wav"
"ecapatdnn_voxceleb12-16k": {
'url':
'https://paddlespeech.bj.bcebos.com/vector/voxceleb/sv0_ecapa_tdnn_voxceleb12_ckpt_0_1_2.tar.gz',
'https://paddlespeech.bj.bcebos.com/vector/voxceleb/sv0_ecapa_tdnn_voxceleb12_ckpt_0_2_0.tar.gz',
'md5':
'cc33023c54ab346cd318408f43fcaf95',
'cfg_path':

@ -181,6 +181,10 @@ int main(int argc, char* argv[]) {
ppspeech::LinearSpectrogramOptions opt;
opt.frame_opts.frame_length_ms = 20;
opt.frame_opts.frame_shift_ms = 10;
opt.frame_opts.dither = 0.0;
opt.frame_opts.remove_dc_offset = false;
opt.frame_opts.window_type = "hanning";
opt.frame_opts.preemph_coeff = 0.0;
LOG(INFO) << "frame length (ms): " << opt.frame_opts.frame_length_ms;
LOG(INFO) << "frame shift (ms): " << opt.frame_opts.frame_shift_ms;

@ -14,6 +14,8 @@
#include "frontend/audio/linear_spectrogram.h"
#include "kaldi/base/kaldi-math.h"
#include "kaldi/feat/feature-common.h"
#include "kaldi/feat/feature-functions.h"
#include "kaldi/matrix/matrix-functions.h"
namespace ppspeech {
@ -21,30 +23,23 @@ namespace ppspeech {
using kaldi::int32;
using kaldi::BaseFloat;
using kaldi::Vector;
using kaldi::SubVector;
using kaldi::VectorBase;
using kaldi::Matrix;
using std::vector;
LinearSpectrogram::LinearSpectrogram(
const LinearSpectrogramOptions& opts,
std::unique_ptr<FrontendInterface> base_extractor) {
opts_ = opts;
std::unique_ptr<FrontendInterface> base_extractor)
: opts_(opts), feature_window_funtion_(opts.frame_opts) {
base_extractor_ = std::move(base_extractor);
int32 window_size = opts.frame_opts.WindowSize();
int32 window_shift = opts.frame_opts.WindowShift();
fft_points_ = window_size;
dim_ = window_size / 2 + 1;
chunk_sample_size_ =
static_cast<int32>(opts.streaming_chunk * opts.frame_opts.samp_freq);
hanning_window_.resize(window_size);
double a = M_2PI / (window_size - 1);
hanning_window_energy_ = 0;
for (int i = 0; i < window_size; ++i) {
hanning_window_[i] = 0.5 - 0.5 * cos(a * i);
hanning_window_energy_ += hanning_window_[i] * hanning_window_[i];
}
dim_ = fft_points_ / 2 + 1; // the dimension is Fs/2 Hz
hanning_window_energy_ = kaldi::VecVec(feature_window_funtion_.window,
feature_window_funtion_.window);
}
void LinearSpectrogram::Accept(const VectorBase<BaseFloat>& inputs) {
@ -56,99 +51,57 @@ bool LinearSpectrogram::Read(Vector<BaseFloat>* feats) {
bool flag = base_extractor_->Read(&input_feats);
if (flag == false || input_feats.Dim() == 0) return false;
vector<BaseFloat> input_feats_vec(input_feats.Dim());
std::memcpy(input_feats_vec.data(),
input_feats.Data(),
input_feats.Dim() * sizeof(BaseFloat));
vector<vector<BaseFloat>> result;
Compute(input_feats_vec, result);
int32 feat_size = 0;
if (result.size() != 0) {
feat_size = result.size() * result[0].size();
}
feats->Resize(feat_size);
// todo refactor (SimleGoat)
for (size_t idx = 0; idx < feat_size; ++idx) {
(*feats)(idx) = result[idx / dim_][idx % dim_];
}
return true;
}
void LinearSpectrogram::Hanning(vector<float>* data) const {
CHECK_GE(data->size(), hanning_window_.size());
for (size_t i = 0; i < hanning_window_.size(); ++i) {
data->at(i) *= hanning_window_[i];
}
}
bool LinearSpectrogram::NumpyFft(vector<BaseFloat>* v,
vector<BaseFloat>* real,
vector<BaseFloat>* img) const {
Vector<BaseFloat> v_tmp;
v_tmp.Resize(v->size());
std::memcpy(v_tmp.Data(), v->data(), sizeof(BaseFloat) * (v->size()));
RealFft(&v_tmp, true);
v->resize(v_tmp.Dim());
std::memcpy(v->data(), v_tmp.Data(), sizeof(BaseFloat) * (v->size()));
real->push_back(v->at(0));
img->push_back(0);
for (int i = 1; i < v->size() / 2; i++) {
real->push_back(v->at(2 * i));
img->push_back(v->at(2 * i + 1));
}
real->push_back(v->at(1));
img->push_back(0);
int32 feat_len = input_feats.Dim();
int32 left_len = reminded_wav_.Dim();
Vector<BaseFloat> waves(feat_len + left_len);
waves.Range(0, left_len).CopyFromVec(reminded_wav_);
waves.Range(left_len, feat_len).CopyFromVec(input_feats);
Compute(waves, feats);
int32 frame_shift = opts_.frame_opts.WindowShift();
int32 num_frames = kaldi::NumFrames(waves.Dim(), opts_.frame_opts);
int32 left_samples = waves.Dim() - frame_shift * num_frames;
reminded_wav_.Resize(left_samples);
reminded_wav_.CopyFromVec(
waves.Range(frame_shift * num_frames, left_samples));
return true;
}
// Compute spectrogram feat
// todo: refactor later (SmileGoat)
bool LinearSpectrogram::Compute(const vector<float>& waves,
vector<vector<float>>& feats) {
int num_samples = waves.size();
const int& frame_length = opts_.frame_opts.WindowSize();
const int& sample_rate = opts_.frame_opts.samp_freq;
const int& frame_shift = opts_.frame_opts.WindowShift();
const int& fft_points = fft_points_;
const float scale = hanning_window_energy_ * sample_rate;
bool LinearSpectrogram::Compute(const Vector<BaseFloat>& waves,
Vector<BaseFloat>* feats) {
int32 num_samples = waves.Dim();
int32 frame_length = opts_.frame_opts.WindowSize();
int32 sample_rate = opts_.frame_opts.samp_freq;
BaseFloat scale = 2.0 / (hanning_window_energy_ * sample_rate);
if (num_samples < frame_length) {
return true;
}
int num_frames = 1 + ((num_samples - frame_length) / frame_shift);
feats.resize(num_frames);
vector<float> fft_real((fft_points_ / 2 + 1), 0);
vector<float> fft_img((fft_points_ / 2 + 1), 0);
vector<float> v(frame_length, 0);
vector<float> power((fft_points / 2 + 1));
for (int i = 0; i < num_frames; ++i) {
vector<float> data(waves.data() + i * frame_shift,
waves.data() + i * frame_shift + frame_length);
Hanning(&data);
fft_img.clear();
fft_real.clear();
v.assign(data.begin(), data.end());
NumpyFft(&v, &fft_real, &fft_img);
feats[i].resize(fft_points / 2 + 1); // the last dimension is Fs/2 Hz
for (int j = 0; j < (fft_points / 2 + 1); ++j) {
power[j] = fft_real[j] * fft_real[j] + fft_img[j] * fft_img[j];
feats[i][j] = power[j];
if (j == 0 || j == feats[0].size() - 1) {
feats[i][j] /= scale;
} else {
feats[i][j] *= (2.0 / scale);
}
// log added eps=1e-14
feats[i][j] = std::log(feats[i][j] + 1e-14);
}
int32 num_frames = kaldi::NumFrames(num_samples, opts_.frame_opts);
feats->Resize(num_frames * dim_);
Vector<BaseFloat> window;
for (int frame_idx = 0; frame_idx < num_frames; ++frame_idx) {
kaldi::ExtractWindow(0,
waves,
frame_idx,
opts_.frame_opts,
feature_window_funtion_,
&window,
NULL);
SubVector<BaseFloat> output_row(feats->Data() + frame_idx * dim_, dim_);
window.Resize(frame_length, kaldi::kCopyData);
RealFft(&window, true);
kaldi::ComputePowerSpectrum(&window);
SubVector<BaseFloat> power_spectrum(window, 0, dim_);
power_spectrum.Scale(scale);
power_spectrum(0) = power_spectrum(0) / 2;
power_spectrum(dim_ - 1) = power_spectrum(dim_ - 1) / 2;
power_spectrum.Add(1e-14);
power_spectrum.ApplyLog();
output_row.CopyFromVec(power_spectrum);
}
return true;
}

@ -49,19 +49,15 @@ class LinearSpectrogram : public FrontendInterface {
virtual void Reset() { base_extractor_->Reset(); }
private:
void Hanning(std::vector<kaldi::BaseFloat>* data) const;
bool Compute(const std::vector<kaldi::BaseFloat>& waves,
std::vector<std::vector<kaldi::BaseFloat>>& feats);
bool NumpyFft(std::vector<kaldi::BaseFloat>* v,
std::vector<kaldi::BaseFloat>* real,
std::vector<kaldi::BaseFloat>* img) const;
bool Compute(const kaldi::Vector<kaldi::BaseFloat>& waves,
kaldi::Vector<kaldi::BaseFloat>* feats);
kaldi::int32 fft_points_;
size_t dim_;
std::vector<kaldi::BaseFloat> hanning_window_;
kaldi::FeatureWindowFunction feature_window_funtion_;
kaldi::BaseFloat hanning_window_energy_;
LinearSpectrogramOptions opts_;
std::unique_ptr<FrontendInterface> base_extractor_;
kaldi::Vector<kaldi::BaseFloat> reminded_wav_;
int chunk_sample_size_;
DISALLOW_COPY_AND_ASSIGN(LinearSpectrogram);
};

Loading…
Cancel
Save