13 KiB
PaddleSpeech is an open-source toolkit on PaddlePaddle platform for a variety of critical tasks in speech, with the state-of-art and influential models.
Speech-to-Text
| Input Audio | Recognition Result |
|---|---|
 |
I knocked at the door on the ancient side of the building. |
|
|
我认为跑步最重要的就是给我带来了身体健康。 |
Text-to-Speech
| Synthetic Audio | |
|---|---|
| Life was like a box of chocolates, you never know what you're gonna get. |
|
| 早上好，今天是2020/10/29，最低温度是-3°C。 |
|
For more synthesized audios, please refer to PaddleSpeech Text-to-Speech samples.
Via the easy-to-use, efficient, flexible and scalable implementation, our vision is to empower both industrial application and academic research, including training, inference & testing modules, and deployment process. To be more specific, this toolkit features at:
- Fast and Light-weight: we provide high-speed and ultra-lightweight models that are convenient for industrial deployment.
- Rule-based Chinese frontend: our frontend contains Text Normalization and Grapheme-to-Phoneme (G2P, including Polyphone and Tone Sandhi). Moreover, we use self-defined linguistic rules to adapt Chinese context.
- Varieties of Functions that Vitalize both Industrial and Academia:
- Implementation of critical audio tasks: this toolkit contains audio functions like Speech Translation, Automatic Speech Recognition, Text-to-Speech Synthesis, Voice Cloning, etc.
- Integration of mainstream models and datasets: the toolkit implements modules that participate in the whole pipeline of the speech tasks, and uses mainstream datasets like LibriSpeech, LJSpeech, AIShell, CSMSC, etc. See also model list for more details.
- Cascaded models application: as an extension of the application of traditional audio tasks, we combine the workflows of aforementioned tasks with other fields like Natural language processing (NLP), like Punctuation Restoration.
Installation
The base environment in this page is
- Ubuntu 16.04
- python>=3.7
- paddlepaddle>=2.2.0
If you want to set up PaddleSpeech in other environment, please see the installation documents for all the alternatives.
Quick Start
Developers can have a try of our model with only a few lines of code.
A tiny DeepSpeech2 Speech-to-Text model training on toy set of LibriSpeech:
cd examples/tiny/s0/
# source the environment
source path.sh
source ../../../utils/parse_options.sh
# prepare data
bash ./local/data.sh
# train model, all `ckpt` under `exp` dir, if you use paddlepaddle-gpu, you can set CUDA_VISIBLE_DEVICES before the train script
./local/train.sh conf/deepspeech2.yaml deepspeech2 offline
# avg n best model to get the test model, in this case, n = 1
avg.sh best exp/deepspeech2/checkpoints 1
# evaluate the test model
./local/test.sh conf/deepspeech2.yaml exp/deepspeech2/checkpoints/avg_1 offline
For Text-To-Speech, try pretrained FastSpeech2 + Parallel WaveGAN on CSMSC:
cd examples/csmsc/tts3
# download the pretrained models and unaip them
wget https://paddlespeech.bj.bcebos.com/Parakeet/pwg_baker_ckpt_0.4.zip
unzip pwg_baker_ckpt_0.4.zip
wget https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip
unzip fastspeech2_nosil_baker_ckpt_0.4.zip
# source the environment
source path.sh
# run end-to-end synthesize
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/synthesize_e2e.py \
--fastspeech2-config=fastspeech2_nosil_baker_ckpt_0.4/default.yaml \
--fastspeech2-checkpoint=fastspeech2_nosil_baker_ckpt_0.4/snapshot_iter_76000.pdz \
--fastspeech2-stat=fastspeech2_nosil_baker_ckpt_0.4/speech_stats.npy \
--pwg-config=pwg_baker_ckpt_0.4/pwg_default.yaml \
--pwg-checkpoint=pwg_baker_ckpt_0.4/pwg_snapshot_iter_400000.pdz \
--pwg-stat=pwg_baker_ckpt_0.4/pwg_stats.npy \
--text=${BIN_DIR}/../sentences.txt \
--output-dir=exp/default/test_e2e \
--inference-dir=exp/default/inference \
--phones-dict=fastspeech2_nosil_baker_ckpt_0.4/phone_id_map.txt
If you want to try more functions like training and tuning, please see Speech-to-Text Quick Start and Text-To-Speech Quick Start.
Model List
PaddleSpeech supports a series of most popular models, summarized in released models with available pretrained models.
Speech-to-Text module contains Acoustic Model and Language Model, with the following details:
| Speech-to-Text Module Type | Dataset | Model Type | Link |
|---|---|---|---|
| Acoustic Model | Aishell | DeepSpeech2 RNN + Conv based Models | deepspeech2-aishell |
| Transformer based Attention Models | u2.transformer.conformer-aishell | ||
| Librispeech | Transformer based Attention Models | deepspeech2-librispeech / transformer.conformer.u2-librispeech / transformer.conformer.u2-kaldi-librispeech | |
| Alignment | THCHS30 | MFA | mfa-thchs30 |
| Language Model | Ngram Language Model | kenlm | |
| TIMIT | Unified Streaming & Non-streaming Two-pass | u2-timit | |
PaddleSpeech Text-to-Speech mainly contains three modules: Text Frontend, Acoustic Model and Vocoder. Acoustic Model and Vocoder models are listed as follow:
| Text-to-Speech Module Type |
Model Type | |
|
|---|---|---|---|
| Text Frontend | tn / g2p | ||
| Acoustic Model | Tacotron2 | LJSpeech | tacotron2-ljspeech |
| TransformerTTS | transformer-ljspeech | ||
| SpeedySpeech | CSMSC | speedyspeech-csmsc | |
| FastSpeech2 | AISHELL-3 / VCTK / LJSpeech / CSMSC | fastspeech2-aishell3 / fastspeech2-vctk / fastspeech2-ljspeech / fastspeech2-csmsc | |
| Vocoder | WaveFlow | LJSpeech | waveflow-ljspeech |
| Parallel WaveGAN | LJSpeech / VCTK / CSMSC | PWGAN-ljspeech / PWGAN-vctk / PWGAN-csmsc | |
| Voice Cloning | GE2E | AISHELL-3, etc. | ge2e |
| GE2E + Tactron2 | AISHELL-3 | ge2e-tactron2-aishell3 | |
Tutorials
Normally, Speech SoTA gives you an overview of the hot academic topics in speech. To focus on the tasks in PaddleSpeech, you will find the following guidelines are helpful to grasp the core ideas.
- Overview
- Quick Start
- Speech-to-Text
- Text-to-Speech
- Released Models
The TTS module is originally called Parakeet, and now merged with DeepSpeech. If you are interested in academic research about this function, please see TTS research overview. Also, this document is a good guideline for the pipeline components.
FAQ and Contributing
You are warmly welcome to submit questions in discussions and bug reports in issues! Also, we highly appreciate if you would like to contribute to this project!
Citation
To cite PaddleSpeech for research, please use the following format.
@misc{ppspeech2021,
title={PaddleSpeech, a toolkit for audio processing based on PaddlePaddle.},
author={PaddlePaddle Authors},
howpublished = {\url{https://github.com/PaddlePaddle/PaddleSpeech}},
year={2021}
}
License and Acknowledge
PaddleSpeech is provided under the Apache-2.0 License.
PaddleSpeech depends on a lot of open source repositories. See references for more information.