Paddle implementation of deepvoice 3 in dynamic graph, a convolutional network based text-to-speech synthesis model. The implementation is based on [Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning](https://arxiv.org/abs/1710.07654).
...
...
@@ -22,7 +22,7 @@ The model consists of an encoder, a decoder and a converter (and a speaker embed
## Project Structure
```text
├── data.py data_processing
├── data.py data_processing
├── ljspeech.yaml (example) configuration file
├── sentences.txt sample sentences
├── synthesis.py script to synthesize waveform from text
...
...
@@ -50,7 +50,7 @@ optional arguments:
The directory to save result.
-g DEVICE, --device DEVICE
device to use
```
```
1.`--config` is the configuration file to use. The provided `ljspeech.yaml` can be used directly. And you can change some values in the configuration file and train the model with a different config.
2.`--data` is the path of the LJSpeech dataset, the extracted folder from the downloaded archive (the folder which contains metadata.txt).
# Copyright (c) 2020 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.
"""
At most cases, we have non-stream dataset, which means we can random access it with __getitem__, and we can get the length of the dataset with __len__.
...
...
@@ -6,10 +19,10 @@ This suffices for a sampler. We implemente sampler as iterable of valid indices.
So the sampler is only responsible for generating valid indices.
"""
importnumpyasnp
importrandom
classSampler(object):
def__init__(self,data_source):
pass
...
...
@@ -23,7 +36,7 @@ class Sampler(object):
classSequentialSampler(Sampler):
def__init__(self,data_source):
self.data_source=data_source
def__iter__(self):
returniter(range(len(self.data_source)))
...
...
@@ -42,12 +55,14 @@ class RandomSampler(Sampler):
"replacement={}".format(self.replacement))
ifself._num_samplesisnotNoneandnotreplacement:
raiseValueError("With replacement=False, num_samples should not be specified, "
"since a random permutation will be performed.")
raiseValueError(
"With replacement=False, num_samples should not be specified, "
@@ -14,9 +14,4 @@ One of the reasons we choose to load data lazily (only load metadata before hand
For deep learning practice, we typically batch examples. So the dataset should comes with a method to batch examples. Assuming the record is implemented as a tuple with several items. When an item is represented as a fix-sized array, to batch them is trivial, just `np.stack` suffices. But for array with dynamic size, padding is needed. We decide to implement a batching method for each item. Then batching a record can be implemented by these methods. For a dataset, a `_batch_examples` should be implemented. But in most cases, you can choose one from `batching.py`.
Paddle fluid implementation of WaveNet, a deep generative model of raw audio waveforms.
WaveNet model is originally proposed in [WaveNet: A Generative Model for Raw Audio](https://arxiv.org/abs/1609.03499).
Our implementation is based on the WaveNet architecture described in [ClariNet: Parallel Wave Generation in End-to-End Text-to-Speech](https://arxiv.org/abs/1807.07281) and can provide various output distributions, including single Gaussian, mixture of Gaussian, and softmax with linearly quantized channels.
Our implementation is based on the WaveNet architecture described in [ClariNet: Parallel Wave Generation in End-to-End Text-to-Speech](https://arxiv.org/abs/1807.07281) and can provide various output distributions, including single Gaussian, mixture of Gaussian, and softmax with linearly quantized channels.
We implement WaveNet model in paddle fluid with dynamic graph, which is convenient for flexible network architectures.
Our model will save model parameters as checkpoints in `./runs/wavenet/${ModelName}/checkpoint/` every 10000 iterations by default.
The saved checkpoint will have the format of `step-${iteration_number}.pdparams` for model parameters and `step-${iteration_number}.pdopt` for optimizer parameters.
The saved checkpoint will have the format of `step-${iteration_number}.pdparams` for model parameters and `step-${iteration_number}.pdopt` for optimizer parameters.
There are three ways to load a checkpoint and resume training (take an example that you want to load a 500000-iteration checkpoint):
1. Use `--checkpoint=./runs/wavenet/${ModelName}/checkpoint/step-500000` to provide a specific path to load. Note that you only need to provide the base name of the parameter file, which is `step-500000`, no extension name `.pdparams` or `.pdopt` is needed.
1. Use `--checkpoint=./runs/wavenet/${ModelName}/checkpoint/step-500000` to provide a specific path to load. Note that you only need to provide the base name of the parameter file, which is `step-500000`, no extension name `.pdparams` or `.pdopt` is needed.
2. Use `--iteration=500000`.
3. If you don't specify either `--checkpoint` or `--iteration`, the model will automatically load the latest checkpoint in `./runs/wavenet/${ModelName}/checkpoint`.
In this example, `--output` specifies where to save the synthesized audios and `--sample` specifies which sample in the valid dataset (a split from the whole LJSpeech dataset, by default contains the first 16 audio samples) to synthesize based on the mel-spectrograms computed from the ground truth sample audio, e.g., `--sample=0` means to synthesize the first audio in the valid dataset.
