Skip to content

Latest commit

 

History

History
247 lines (197 loc) · 7.97 KB

README.md

File metadata and controls

247 lines (197 loc) · 7.97 KB

Text2Video

This is code for ICASSP 2022: "Text2Video: Text-driven Talking-head Video Synthesis with Phonetic Dictionary". Project Page

ICASSP 2022: Text2Video

Introduction

With the advance of deep learning technology, automatic video generation from audio or text has become an emerging and promising research topic. In this paper, we present a novel approach to synthesize video from text. The method builds a phoneme-pose dictionary and trains a generative adversarial network (GAN) to generate video from interpolated phoneme poses. Compared to audio-driven video generation algorithms, our approach has a number of advantages: 1) It only needs a fraction of the training data used by an audio-driven approach; 2) It is more flexible and not subject to vulnerability due to speaker variation; 3) It significantly reduces the preprocessing, training and inference time. We perform extensive experiments to compare the proposed method with state-of-the-art talking face generation methods on a benchmark dataset and datasets of our own. The results demonstrate the effectiveness and superiority of our approach.

Data / Preprocessing

Set up

  1. Git clone repo
git clone git@github.com:sibozhang/Text2Video.git
  1. Download and install modified vid2vid repo vid2vid

  2. Download Trained model

Please build 'checkpoints' folder in vid2vid folder and put trained model in it.

VidTIMIT fadg0 (English, Female)

Dropbox: https://www.dropbox.com/sh/lk6et49v2uyfzjx/AADAFAp02_b3FQchaYxOZ0EMa?dl=0

百度云链接: https://pan.baidu.com/s/1SSkMKOK9LhClW2JvDCSiLg?pwd=bevj 提取码: bevj

Xuesong (Chinese, Male)

Dropbox: https://www.dropbox.com/sh/qz3zoma5ac9mw5p/AAARiR8xKvATN4CBSyjWt_uOa?dl=0

百度云链接: 链接: https://pan.baidu.com/s/1DvuBbThYo4n5RIZsc-92rg?pwd=am7d 提取码: am7d

  1. Prepare data and folder in the following order

    Text2Video
    ├── *phoneme_data
    ├── model
    ├── ...
    vid2vid
    ├── ...
    venv
    ├── vid2vid
    
  2. Setup env

sudo apt-get install sox libsox-fmt-mp3
pip install zhon
pip install moviepy
pip install ffmpeg
pip install dominate
pip install pydub

For Chinese, we use vosk to get timestamp of each words. Please install vosk from https://alphacephei.com/vosk/install and unpack as 'model' in the current folder. or install:

pip install vosk
pip install cn2an
pip install pypinyin

Testing

  1. Activate vitrual environment vid2vid
source ../venv/vid2vid/bin/activate
  1. Generate video with real audio in English
sh text2video_audio.sh $1 $2

Generate video with TTS audio in English

sh text2video_tts.sh $1 $2 $3

Generate video with TTS audio in Chinese

sh text2video_tts.sh $1 $2 $3

$1: "input text" $2: person $3: fill f for female or m for male (gender)

Example 1. test VidTIMIT data with real audio.

sh text2video_audio.sh "She had your dark suit in greasy wash water all year." fadg0 f

Example 2. test VidTIMIT data with TTS audio.

sh text2video_tts.sh "She had your dark suit in greasy wash water all year." fadg0 f

Example 3. test with Chinese female TTS audio.

sh text2video_tts_chinese.sh "正在为您查询合肥的天气情况。今天是2020年2月24日,合肥市今天多云,最低温度9摄氏度,最高温度15摄氏度,微风。" henan f

Training with your own data

English Phoneme / Chinese Pinyin model:

  1. Modeling

1.1 Video recording:

Read prompts cover all phonemes or pinyin, please refer to prompts.docx for phoneme or all_pinyin.txt for pinyin under ./prompts folder. Pause 0.5 second between each pronunciation. Use the camera to record at least 1280x720 video.

1.2 Phoneme-Mouth/ Pinyin-Mouth Shape Dictionary:

Use montreal-forced-aligner (google STT) for Phoneme/ vosk for pinyi to get timestamp of each words and put it in a dictionary file, which stores the data structure is as follows, each line saves a pair of [phoneme/ pinyin, frame number]:

phoneme, frame: AA 52 AA0 52 AA1 52 AA2 52 AE 90 AE0 90 AE1 90 AE2 90 AH 127 AH0 127 AH1 127 AH2 127 AO 146 AO0 146 AO1 146 AO2 146 AW 227 AW0 227 AW1 227 AW2 227 ...

pinyin, frame: ba 61 bo 86 bi 540 bu 110 bai 130 bao 154 ban 178 bang 202 ou 225 pa 272 po 298 ...

1.3 Openpose:

Use openpose to fit each frame of the video, find out the skeleton result of the human body, and save it to a separate folder. The code of openpose can be downloaded at: https://github.com/CMU-Perceptual-Computing-Lab/openpose After the compilation is successful, run:

./build/examples/openpose/openpose.bin --image_dir ./images --face --hand --write_json ./keypoints

1.4 Use the generated human skeleton model and the corresponding video to train the vid2vid model, which is used to generate a realistic human skeleton model portrait video. The code for vid2vid can be downloaded from here: https://github.com/NVIDIA/vid2vid

Train:

python train.py --name xx --dataroot datasets/xx --dataset_mode pose --input_nc 3
--openpose_only --num_D 2 --resize_or_crop randomScaleHeight_and_scaledCrop
--loadSize 544 --fineSize 512 --gpu_ids 0,1,2,3,4,5,6,7 --batchSize 8
--max_frames_per_gpu 2 --niter 500 --niter_decay 5 --no_first_img --n_frames_total 12
--max_t_step 4 --niter_step 100 --save_epoch_freq 100 --add_face_disc
--random_drop_prob 0
  1. Video generation

2.1 Generate audio files from text:

Use Baidu TTS cloud service to generate the required voice. Baidu voice cloud service address: http://wiki.baidu.com/pages/viewpage.action?pageId=342334101. Please refer to the specific call tts_request.py

2.2 Analyze the audio and find out the time stamp of each text:

use the Chinese version of VOSK speech recognition to process the speech generated by TTS Line recognition, and generate the following <frame number, pinyin> file. The VOSK code can be downloaded here: https://github.com/alphacep/vosk-api. For specific calls, please refer to pinyin_timestamping.py

25 xu 29 yao 38 zuo 46 Geng 53 jia 60 chong 65 fen 70 de 75 zun 80 bei 100 yin 105 ci 111 ne 116 wei 118 le

2.3 For each text in the audio, use its pinyin to find the corresponding 2D skeleton model, and splice it into a dynamic 2D skeleton. For video, intermediate frames are obtained by interpolation. Please refer to interp_landmarks_motion.py for details.

2.4 Use the vid2vid model to generate the final portrait video from the 2D skeleton video. The example is as follows:

CUDA_VISIBLE_DEVICES=1 python test.py --name xx --dataroot datasets/xx
--dataset_mode pose --input_nc 3 --resize_or_crop scaleHeight --loadSize 512
--openpose_only --how_many 1200 --no_first_img --random_drop_prob 0

Citation

Please cite our paper in your publications.

Sibo Zhang, Jiahong Yuan, Miao Liao, Liangjun Zhang. PDF Result Video

@INPROCEEDINGS{9747380,  
author={Zhang, Sibo and Yuan, Jiahong and Liao, Miao and Zhang, Liangjun},  
booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},   
title={Text2video: Text-Driven Talking-Head Video Synthesis with Personalized Phoneme - Pose Dictionary},   
year={2022},  
volume={},  
number={},  
pages={2659-2663},  
doi={10.1109/ICASSP43922.2022.9747380}
}
@article{zhang2021text2video,
  title={Text2Video: Text-driven Talking-head Video Synthesis with Phonetic Dictionary},
  author={Zhang, Sibo and Yuan, Jiahong and Liao, Miao and Zhang, Liangjun},
  journal={arXiv preprint arXiv:2104.14631},
  year={2021}
}

Appendices

ARPABET

Ackowledgements

This code is based on the vid2vid framework.